# Reason — full content export > Every canonical Reason Machines public page concatenated as Markdown for single-fetch ingestion. > Canonical index: https://reasonmachines.company/llms.txt > Source: https://reasonmachines.company/ # Reason Machines is building the world’s best coding agent. In the new era of harnesses, moving fast on the best way to utilize new models and their capabilities is everything. Our principle is minimal and maximal harnesses: minimal so we stay the fastest to act, and maximal so we use frontier models for their strengths. New: [Reason Agent 1.0 on DeepSWE](/blog/reason-agent-deepswe) [contact@reasonmachines.company](mailto:contact@reasonmachines.company) Reason is an autonomous software engineer for GitHub. Connect a repository, assign issues, and Reason's cloud agents open evidence-backed pull requests with screen recordings, logs, and review-ready diffs. People drive Reason from the web app at https://reasonmachines.company/agents; programmatic per-user access is available through the HTTP API documented at https://reasonmachines.company/llms.txt. ## Pages Every page is also available as clean Markdown: append `.md` to any URL, or send `Accept: text/markdown`. The entire site as one file: https://reasonmachines.company/llms-full.txt ### Blog - [Reason for macOS](https://reasonmachines.company/blog/ara-for-macos.md): Run cloud sessions and let Reason work with your local files, tools, and ports. - [Introducing Reason Agent 1.0](https://reasonmachines.company/blog/reason-agent-deepswe.md): Our first results on DeepSWE v1.1. Competitive with frontier harnesses, with up to 39% lower median solving time against our benchmark baseline. - [Rust harness wars: grok-build vs codex](https://reasonmachines.company/blog/rust-harness-wars.md): We cloned xAI's grok-build and OpenAI's codex and read both line by line, twenty subsystems each, to see where two labs building the same coding agent actually diverge. - [Reason for Startups](https://reasonmachines.company/blog/yc.md): Current and alumni YC companies get $1,000 in credits to start coding. Open the Reason deal on Bookface and follow the redemption instructions to claim them. - [Building the software factory](https://reasonmachines.company/blog/about.md): We are building the software factory we wanted as engineers: fast enough to keep up with ideas, careful enough to trust with real repositories. ### Legal - [Terms of Service](https://reasonmachines.company/terms-of-service.md): Terms governing Reason Machines websites, applications, APIs, cloud sessions, and local-runtime controls. - [Privacy Policy](https://reasonmachines.company/privacy-policy.md): How Reason Machines collects, uses, shares, retains, and deletes information. - [Account Deletion](https://reasonmachines.company/delete-account.md): How to request deletion of your Reason account and associated data. --- > Source: https://reasonmachines.company/blog/ara-for-macos # Reason for macOS _Run cloud sessions and let Reason work with your local files, tools, and ports._ By Reason Machines August 29, 2026 Reason for macOS connects a Reason session to your Mac. A task can use a local repository, installed tools, files, and development ports when you allow it. Use the desktop app when the code or service Reason needs is already running on your computer. ## Work with the local environment Reason can read the selected checkout, run installed tools, and open local development ports. Local access is granted per task. ![Reason for macOS showing a coding session, local computer controls, and a repository review side by side](/product/ara-macos-local-environment.webp) ## Review and control the run See active permissions, tool activity, and changed files. Send a follow-up or stop the run from the same window. ## Use the same session elsewhere Open the same session on the web or iOS. Its messages and results remain attached. --- > Source: https://reasonmachines.company/blog/reason-agent-deepswe # Introducing Reason Agent 1.0 _Our first results on DeepSWE v1.1. Competitive with frontier harnesses, with up to 39% lower median solving time against our benchmark baseline._ By Reason Machines September 14, 2026 [Reason Agent](/) is the autonomous software engineer built by Reason Machines. Today we’re introducing Reason Agent 1.0, our coding-agent harness built to help models solve software engineering tasks. Our first results on DeepSWE v1.1 are competitive with frontier harnesses, with up to 39% lower median solving time against our Codex benchmark baseline.[1](#ref-1) ### DeepSWE score vs. median wall time DeepSWE score Higher score · less time ↗Configs (60/71) gpt-6-astraxhighhighmaxmediumlowgemini-3.8-flashhighmediumclaude-opus-5maxxhighhighmediumlowgpt-5.6-solmaxxhighhighmediumlowclaude-fable-5xhighmaxhighmediumlowgpt-5.6-terramaxxhighhighmediumlowglm-5.3maxkimi-k3maxgrok-4.6mediumxhighhighlowgpt-5.6-lunamaxxhighhighmediumlowgpt-5.5xhighhighmediumlowgemini-3.7-flashmediumhighlowglm-5.3-flashmaxdeepseek-v4-promaxclaude-opus-4.8maxxhighhighmediumlowqwen3.8-maxxhighmuse-spark-1.2xhighclaude-sonnet-5maxxhighhighmediumlowgrok-4.5highdeepseek-v4-flashmaxmuse-spark-1.1xhighgpt-5.4xhighgemini-3.6-flashhighglm-5.2maxhighgemini-3.5-flashhighkimi-k2.7-codedefaultclaude-sonnet-4.6highgemini-3.1-pro-previewhighReason AgentAstra xhigh Reason AgentAstra · xhigh · our runs 73.0%DeepSWE score 13.80 minMedian wall time Figure 1. Median recorded agent-execution wall time, excluding setup, queueing and verification, not end-to-end task latency. Published medians cover scored attempts, including failures; Reason pools timed attempts across both runs and retains all assignments in its score. Model lines connect configurations, not a fitted frontier. Different hosts, provider load and execution conditions make this descriptive, not a matched speed test. [[1]](#ref-1) [[2]](#ref-2) ## Successful solutions, earlier Reason is our coding-agent harness, built around a minimal set of general-purpose tools. The model decides how to inspect a repository, make changes and test its work. We give it room to adapt its approach to the problem, rather than prescribe a fixed sequence of task-specific procedures. We reject the assumption that better agents necessarily need more orchestration. As models become more capable, our view is that the harness should impose less workflow and place more trust in the model’s problem-solving decisions. Its responsibility is to provide reliable tools and clear execution boundaries, without deciding every step on the model’s behalf. The aim is to reduce harness-generated context and forced coordination calls, giving the model a more direct path from investigation to a tested change. That is a design hypothesis, not a mechanism established by these runs. A small toolset alone does not guarantee fewer model turns. ## Alongside published reference points Reason’s Astra result sits close to Datacurve’s leading published Astra configuration. The table combines our measurements with the public reference points.[2](#ref-2)[3](#ref-3)[1](#ref-1) ### Reason Agent is competitive with leading harnesses DeepSWE v1.1 · observed success Agent | Effort | Success | Reason (Astra) | xhigh | 73%67–78% | GPT-6 Astra | xhigh | 74%70–78% | Gemini 3.8 Flash | high | 74%70–78% | Claude Opus 5 | max | 74%69–78% | GPT-5.6 Sol | max | 73%68–77% | Claude Fable 5 | xhigh | 70%66–74% | Figure 2. Reason (Astra) is our measurement; all other rows use mini-swe-agent and published DeepSWE results. Ranges show 95% Wilson intervals, unadjusted for repeated-task dependence. Evaluation methods differ. [[1]](#ref-1)[[2]](#ref-2)[[3]](#ref-3)[[6]](#ref-6) ## A side-by-side comparison of harnesses Same model and effort, different harnesses. Reason used 36.7% less mean solving time on shared successes, averaging its attempts per task. Its overall median solving time was about 40% lower (39% across the combined runs) than Codex’s.[1](#ref-1) We compare Reason and Codex on the same 113 tasks, both using Astra at xhigh effort. The curve combines both Reason runs and the single Codex run. Below it, the task-by-task view compares run 1 from each harness, including the two Reason recoveries, so each dot shows how both performed on the same task. The completion curve shows the measured result: successful attempts finish earlier, at a similar overall observed success rate. By 20 minutes of solving time, final passing attempts accounted for 59.3% of Reason assignments versus 20.4% for Codex. These are retrospective completion times, not results from an experiment with a 20-minute deadline.[1](#ref-1) ### Reason and Codex: solving time Astra xhigh · two Reason runs, one Codex run Click a legend item to show or hide it. Confirmed successes (%) 0 20 40 60 80 100 0102030405060 Solving time (minutes) Figure 3. Final passing attempts by recorded agent time, not an early-stopping experiment. Reason and Codex retain all assignments in their denominators. Both use Astra at xhigh effort; separate provider routes and execution conditions limit causal claims about the harness. [[1]](#ref-1) ### Reason and Codex: task by task Astra xhigh · run 1 from each harness · one dot per task Left half: Reason · Right half: Codex abs-module-cache-flags. Reason: pass; Codex: pass abs-stepped-slices. Reason: pass; Codex: pass actionlint-action-pinning-lint. Reason: pass; Codex: pass adaptix-name-mapping-aliases. Reason: pass; Codex: pass aiomonitor-task-snapshots-diff. Reason: pass; Codex: pass anko-default-function-arguments. Reason: pass; Codex: pass anko-typed-variable-bindings. Reason: pass; Codex: pass arcane-drift-detection-baselines. Reason: pass; Codex: fail arktype-json-schema-refs-dependencies. Reason: pass; Codex: pass awilix-async-container-initialization. Reason: pass; Codex: fail bandit-incremental-cache-control. Reason: pass; Codex: pass bandit-interprocedural-taint-checks. Reason: pass; Codex: pass bandit-structured-nosec-directives. Reason: fail; Codex: fail boa-hierarchical-evaluation-cancellation. Reason: pass; Codex: pass cattrs-partial-structuring-recovery. Reason: pass; Codex: pass clack-async-autocomplete-options. Reason: pass; Codex: fail claude-code-by-agents-recursive-delegation. Reason: pass; Codex: pass cliffy-config-file-parsing. Reason: pass; Codex: pass csstree-shorthand-expansion-compression. Reason: pass; Codex: pass dasel-html-document-format. Reason: pass; Codex: pass dateutil-rfc5545-timezone-interop. Reason: pass; Codex: pass drizzle-orm-window-function-builders. Reason: pass; Codex: pass dynamodb-toolbox-conditional-attribute-requirements. Reason: pass; Codex: pass dynamodb-toolbox-lazy-recursive-schemas. Reason: pass; Codex: pass effect-sse-httpapi-streaming. Reason: pass; Codex: fail eicrud-keyset-pagination-cursor. Reason: pass; Codex: pass etree-xml-diff-patch. Reason: fail; Codex: fail expr-try-catch-errors. Reason: pass; Codex: pass fastapi-deprecation-response-headers. Reason: pass; Codex: pass fastapi-implicit-head-options. Reason: pass; Codex: pass fd-deterministic-multi-key-sorting. Reason: pass; Codex: pass geo-shapeindex-serialization. Reason: pass; Codex: pass go-critic-doc-link-checker. Reason: fail; Codex: fail go-genai-streamed-function-args. Reason: pass; Codex: pass go-git-worktree-merge-conflicts. Reason: pass; Codex: pass goreleaser-retry-publish-auditing. Reason: pass; Codex: pass gql-incremental-graphql-delivery. Reason: fail; Codex: fail happy-dom-abort-pending-body-reads. Reason: pass; Codex: pass happy-dom-deterministic-intersectionobserver. Reason: fail; Codex: fail helm-array-merge-strategies. Reason: fail; Codex: fail helm-unified-manifest-stream. Reason: pass; Codex: pass httpx-deterministic-cookie-store. Reason: pass; Codex: pass httpx-multipart-response-parsing. Reason: pass; Codex: pass httpx-streaming-json-iteration. Reason: pass; Codex: pass igel-persist-feature-schema. Reason: fail; Codex: fail ink-grid-box-layout. Reason: fail; Codex: fail ipython-session-bundle-replay. Reason: pass; Codex: pass katex-multicolumn-array-spans. Reason: pass; Codex: pass kcp-go-multiplexed-kcp-streams. Reason: pass; Codex: fail kea-atomic-signal-selectors. Reason: fail; Codex: fail kgateway-consistent-hash-policy. Reason: pass; Codex: pass kombu-single-active-consumer-priority. Reason: pass; Codex: pass kombu-virtual-queue-dead-lettering. Reason: pass; Codex: pass koota-composite-trait-aspects. Reason: pass; Codex: pass koota-deferred-mutation-buffer. Reason: fail; Codex: pass koota-entity-snapshot-rollback. Reason: pass; Codex: fail koota-pair-relation-tracking. Reason: fail; Codex: fail koota-query-predicates. Reason: pass; Codex: fail kysely-window-grouping-helpers. Reason: pass; Codex: pass langchain-request-coalescing. Reason: pass; Codex: pass mashumaro-flattened-dataclass-fields. Reason: pass; Codex: pass meriyah-explicit-resource-declarations. Reason: fail; Codex: fail mnamer-daemon-watch-lifecycle. Reason: pass; Codex: pass mobly-grouped-test-barriers. Reason: pass; Codex: pass narwhals-rolling-window-suite. Reason: pass; Codex: pass numba-stencil-boundary-modes. Reason: pass; Codex: pass obsidian-linter-auto-table-of-contents. Reason: fail; Codex: fail obsidian-linter-link-format-conversion. Reason: fail; Codex: fail obsidian-linter-scoped-ignore-markers. Reason: pass; Codex: pass ofetch-per-origin-circuit-breaker. Reason: pass; Codex: pass onedump-dump-encryption-pipeline. Reason: fail; Codex: fail opa-rego-rule-profiling. Reason: fail; Codex: fail opa-template-string-reconstruction. Reason: pass; Codex: fail optique-conditional-option-dependencies. Reason: fail; Codex: pass oxvg-structural-selector-preservation. Reason: fail; Codex: pass participle-grammar-conflict-analysis. Reason: fail; Codex: fail pebble-durability-wait-apis. Reason: pass; Codex: pass pest-character-class-coalescing. Reason: fail; Codex: fail prometheus-transactional-reload-status. Reason: fail; Codex: pass prometheus-typed-label-sorting. Reason: pass; Codex: pass psd-tools-blend-range-api. Reason: pass; Codex: pass pwntools-tube-multiplexing. Reason: pass; Codex: fail python-statemachine-state-data-scoping. Reason: fail; Codex: fail query-persist-restored-query-state. Reason: pass; Codex: pass quill-shared-toolbar-focus. Reason: pass; Codex: pass returns-validated-error-accumulation. Reason: pass; Codex: pass scc-bounded-memory-spilling. Reason: pass; Codex: pass scriggo-method-declarations. Reason: pass; Codex: pass skrub-duration-encoding. Reason: pass; Codex: pass sql-formatter-bigquery-pipe-formatting. Reason: pass; Codex: pass sqlfmt-create-table-ddl-formatting. Reason: fail; Codex: fail sqlite-utils-safe-import-checkpoints. Reason: pass; Codex: pass superjson-error-stack-serialization. Reason: pass; Codex: pass task-task-graph-export. Reason: pass; Codex: pass tengo-callable-instance-isolation. Reason: pass; Codex: pass tengo-destructuring-bindings. Reason: pass; Codex: pass termenv-preserve-ansi-resets. Reason: fail; Codex: fail testem-bail-on-test-failure. Reason: fail; Codex: fail testem-per-launcher-reports. Reason: fail; Codex: pass textual-kitty-key-phases. Reason: pass; Codex: pass textual-richlog-follow-state. Reason: pass; Codex: pass tomlkit-toml-table-converters. Reason: pass; Codex: pass true-myth-iterable-collection-combinators. Reason: pass; Codex: pass ts-pattern-match-each. Reason: fail; Codex: fail updo-policy-alerting. Reason: pass; Codex: pass valibot-recursive-schema-composition. Reason: pass; Codex: pass vitest-duration-sharding. Reason: pass; Codex: pass vulture-persistent-analysis-cache. Reason: fail; Codex: fail wasmi-trap-coredumps. Reason: pass; Codex: pass wazero-multi-module-snapshots. Reason: fail; Codex: fail yaegi-go-embed-directives. Reason: pass; Codex: pass yjs-map-conflict-detection. Reason: pass; Codex: pass ytt-jsonpath-query-api. Reason: pass; Codex: pass Both passReason passesCodex passesNeither Figure 4. Reason run 1 versus Codex, including the Kea setup-failure retry and unchanged-patch Pwntools regrade. Blue means a pass. Hover or focus a dot to inspect both outcomes. [[1]](#ref-1) ## What comes next These results are preliminary. We believe coding agents should be built around a minimal set of general-purpose tools, giving capable models more freedom to solve tasks. Next, we will test that principle on fresh tasks under matched conditions, varying tools and orchestration one at a time to see whether less context and fewer calls can deliver faster results without sacrificing correctness. ## Evaluation notes - These measurements evaluate the extracted coding harness, not the hosted product. Separate provider routes and execution infrastructure limit causal and formal-parity claims.[1](#ref-1) - Solving time excludes setup, pre-trial queueing and verification. Figure 4 shows Reason run 1 only; aggregate figures retain both runs.[1](#ref-1) - Updated September 14: Reason has 165 passes, 61 failures and no ungraded outcomes across 226 assignments. Kea failed on a fresh attempt after its setup failure; Pwntools passed when its unchanged patch was regraded. No graded failures were replaced. The JSON preserves original outcomes and recovery details. Published scores retain Datacurve’s scoring rules.[1](#ref-1)[2](#ref-2)[3](#ref-3) - Run results, aggregate scores and solving times are in the results JSON.[1](#ref-1) ## References [1] Reason Machines. DeepSWE v1.1. [Run results and aggregate data (JSON)](/research/reason-agent-deepswe/deepswe.json). [2] Datacurve. DeepSWE v1.1. [Reporting methodology](https://deepswe.datacurve.ai/blog/deepswe-v1-1) · [published leaderboard](https://deepswe.datacurve.ai/artifacts/v1.1/leaderboard-live.json). [3] Datacurve. DeepSWE v1.1 rollout data. [Trial browser](https://deepswe.datacurve.ai/data/v1.1) · [raw trial index](https://deepswe.datacurve.ai/artifacts/v1.1/trials.json). [4] Datacurve. DeepSWE: Measuring frontier coding agents. [Evaluation methodology](https://deepswe.datacurve.ai/blog/deepswe#evaluation-harness). [5] Datacurve / Harbor. DeepSWE v1.1, 113 tasks. [Dataset](https://hub.harborframework.com/datasets/datacurve/deep-swe-1-1/latest). [6] NIST/SEMATECH. e-Handbook of Statistical Methods, §7.2.4.1. [Wilson confidence intervals for a proportion](https://www.itl.nist.gov/div898/handbook/prc/section2/prc241.htm). --- > Source: https://reasonmachines.company/blog/rust-harness-wars # Rust harness wars: grok-build vs codex _We cloned xAI's grok-build and OpenAI's codex and read both line by line, twenty subsystems each, to see where two labs building the same coding agent actually diverge._ By Reason Machines July 23, 2026 Two of the most capable terminal coding agents in the world are written in Rust, ship as a single binary, and were open sourced within months of each other: xAI's [grok-build](https://github.com/xai-org/grok-build), the grok CLI, and OpenAI's [codex](https://github.com/openai/codex). We cloned both and read them line by line, roughly a hundred analysis agents across twenty subsystems, to answer one question. When two labs build the same tool, where do their instincts actually diverge? The uncomfortable starting point is in grok-build's own [THIRD-PARTY-NOTICES](https://github.com/xai-org/grok-build/blob/main/THIRD-PARTY-NOTICES). It ports tool implementations from [openai/codex](https://github.com/openai/codex) and [sst/opencode](https://github.com/sst/opencode). This is not a clean room rivalry. It is a descendant remixing its ancestor, which makes the places they differ all the more revealing. ## One skeleton, two instincts Read side by side, the two agents share a near identical skeleton. A single session actor sits behind a typed message boundary. One binary wears many thin faces: an interactive terminal UI, a headless exec mode, and an editor server. The core loop is the same, sample the model, run the tools it asks for, repeat until it stops. Both persist an append only log as the durable source of truth, both compact the conversation when the context window fills, both can resume and fork a session. So the interesting part is not the shape. It is what each team refused to compromise on once they had it. Codex subtracts. grok-build reaches. ## Codex subtracts Codex is the work of a team that owns its model and optimizes for control, safety, and a small surface. Its instinct is to delete optionality. It removed the older chat wire protocol so there is exactly one path to the model. It dropped temperature and top-p sampling in favor of a single reasoning effort knob. It refuses to be a provider marketplace, shipping only a short list of first party backends. That focus buys depth on the surfaces that actually bite. A three platform sandbox built from source for determinism. A policy engine with per turn, feature gated tool planning. An approval and escalation layer wrapped around every command the model runs. A review enforced rule that the model visible history is sacred and is never rewritten. Codex reads like a platform team building for a fleet of machines it will never see. ## grok-build reaches grok-build is more architecturally ambitious and, in places, less finished. It deleted nothing and reached further. A machine wide leader process serves many warm clients over the [Agent Client Protocol](https://agentclientprotocol.com), the emerging cross vendor standard, used here as the internal wire rather than a bolt on. A provider agnostic sampler fans out to three different model protocols, so adding a provider is config, not code. Memory is a real subsystem, markdown files indexed for retrieval rather than a grep over past turns. It ships a [tree-sitter](https://tree-sitter.github.io/tree-sitter/) code graph for structural navigation, and session memory that compounds across runs. This is the topology of a tool designed to live inside other people's environments and outlast any single model. grok-build hedges against lock in the same way codex hedges against untrusted machines. ## The scorecard Across the twenty subsystems we compared, codex holds the stronger design on eleven, the two are a genuine toss up on eight, and one is a tie. grok-build does not win an axis outright, and the reason is honest. codex is a mature, upstream product. grok-build is a v0.1 alpha, and it shows. But the eight toss ups are where the ambition pays off. Broader model support, memory retrieval, the editor protocol, cloud execution, and exact cost accounting are all places where reaching further produced a genuinely different, sometimes better, answer. A tally that reads eleven to nothing on paper is really a story about maturity racing ambition. 11 Codex Advantage8 grok-build Toss-ups1 Direct Tie Subsystem | Verdict | Architectural Analysis | Sandboxing & Isolation | Codex | Codex ships 3-OS source-built sandbox (macOS/Linux/Windows). grok-build covers 2 OS. | Safety & Policy Gates | Codex | Codex enforces per-turn feature gates and escalation prompts. grok-build safety hooks fail open. | Model Protocol Multiplexing | Toss-Up | grok-build cleanly dispatches between OpenAI, Anthropic, and xAI. Codex locks to first-party wire. | Agent Client Protocol (ACP) | Toss-Up | grok-build adopts emerging standard ACP for cross-vendor client-daemon architecture. | Memory & State Compounding | Toss-Up | grok-build indexes markdown files for retrieval. Codex replays and compacts raw transcripts. | Code Graph & Tree-Sitter | Toss-Up | grok-build includes native Tree-sitter AST navigation crate for repository structural maps. | Skill Interoperability | Toss-Up | grok-build consumes Claude and Cursor SKILL.md folders without manual conversion. | Skill Depth & Capability | Codex | Codex skills are full executable packages with unit test scripts and UI metadata. | Tool Governance Middleware | Codex | Codex wraps execution in strict platform policy guards and validation middlewares. | Transcript Integrity | Codex | Codex enforces immutable, unrewritten model history for reproducible audits. | Configuration Hardening | Codex | Codex embeds cryptographic signing keys. grok-build signing is inert in v0.1. | Headless Execution Engine | Codex | Codex headless CI runner is deterministic and battle-tested. | Cost & Token Accounting | Toss-Up | grok-build maintains precise per-turn token and dollar ledgers across mixed backends. | Editor Server Integration | Toss-Up | grok-build integrates natively as an ACP language server. | Error Recovery & Continuation | Codex | Codex has sophisticated loop detection and automatic circuit breaking. | Terminal UI (TUI) Safety | Codex | Codex TUI is 100% safe Rust. grok-build uses raw pointer casts. | Release Packaging & Portability | Tie | Both compile into single static standalone Rust binaries with zero external runtimes. | Prompt Pruning & Compaction | Codex | Codex aggressive context compaction preserves reasoning trajectory. | Cloud Execution Wire | Toss-Up | grok-build daemon topology enables remote headless worker handoffs. | Telemetry & Observability | Codex | Codex structured event traces enable end-to-end replay verification. | ## Skills, memory, and tools On skills, both adopt the same [SKILL.md](https://github.com/anthropics/anthropic-quickstarts/blob/main/computer-use-demo/skills.md) convention with progressive disclosure, only a name and description in context until a skill is actually opened. grok-build optimizes for interoperability, reading Claude and Cursor skill directories unmodified, so an existing library just works. codex optimizes for depth, where a skill is a folder with scripts, references, and UI metadata, a richer unit of capability. On memory, codex keeps the raw transcript and rebuilds from it. grok-build runs a dedicated crate that persists human editable markdown and indexes it for retrieval. On tools, codex splits a tool into a model visible spec and an execution runtime with a whole governance middleware on top. grok-build models tools as a strict three crate contract, cleanly separated but lighter on the safety layer. The same pattern repeats everywhere. codex builds the deeper machine, grok-build builds the more portable one. ## The alpha tax grok-build's reach comes with a bill, and reading the source makes it concrete. A configuration signing subsystem ships with two thousand lines of tests but no embedded public keys, so it is effectively inert. Hooks fail open, which means a crashed safety hook silently stops enforcing. The sandbox covers two platforms where codex covers three. There is an unsafe pointer cast in the terminal UI guarded by hand rather than by the type system. None of this is fatal for an alpha, and most of it is the ordinary signature of a young project whose threat model thinking has outrun its shipped hardening. But it is the price of building the more ambitious thing before the more finished one. Inert Config SigningSecurity Deficit 2,000 lines of test fixtures with zero embedded public root keys in shipped binaries. Signature enforcement passes vacuously. Fail-Open Safety HooksPolicy Risk If an external security pre-hook panics or crashes, execution continues unchecked instead of aborting the turn. 2-OS vs 3-OS SandboxingCoverage Gap Covers macOS and Linux only. Windows developer environments execute unsandboxed (Codex provides AppContainer isolation). Unsafe TUI Pointer CastRust Invariant Manual pointer casting in terminal event handling loop, relying on developer discipline rather than the borrow checker. ## Why we care Reason is a company built around agent harnesses, so reading two of the best in the world from the inside is not idle curiosity. The lesson is the one we keep relearning. The model is a commodity you rent. The harness is the product you own. Codex and grok-build are two credible bets on what that harness should be, one subtractive and safe, one additive and portable, and the gap between them is the whole game. ## How this was built Both repositories were cloned and read from source, not from documentation or memory. A multi agent workflow fanned out roughly a hundred analysis agents: architecture mappers, forty subsystem readers running one per repo across twenty dimensions, twenty head to head comparators, and a synthesis panel, each returning a schema validated analysis. Maturity and edge calls are judgments from the code as read, not vendor claims. grok-build at v0.1.220 alpha, codex at current main. --- > Source: https://reasonmachines.company/blog/yc # Reason for Startups _Current and alumni YC companies get $1,000 in credits to start coding. Open the Reason deal on Bookface and follow the redemption instructions to claim them._ By Reason Machines September 3, 2026 The batch is for talking to users, finding product-market fit, and moving fast—not losing full days to mechanical bug fixes, boilerplate chores, and dependency rot. Current and alumni YC founders can redeem exclusive Reason credits directly through Bookface to accelerate their engineering loop. Reason acts as an autonomous software engineer on your team, handling the repetitive coding work and bringing verified pull requests back for your review. ## Run the company, let Reason handle the code Whether you need to reproduce an intermittent checkout bug, trace an authentication webhook, or ship an integration endpoint, Reason executes the task autonomously in an isolated cloud environment and returns a review-ready PR with passing tests. ![Reason working on a software task and reviewing code changes](/product/reason-work.webp) ## Monitor what matters Focus on the metrics that define your company—MRR, user retention, activation rates, and customer conversations—while Reason handles the codebase maintenance in parallel. ![Reason model usage and activity overview](/product/reason-security.webp) ## How to redeem Active batch founders and YC alumni can claim their deal by navigating to the [Reason deal on Bookface](https://bookface.ycombinator.com/deals/6551) or signing in with your company email at [reasonmachines.company/login](/login). --- > Source: https://reasonmachines.company/blog/about # Building the software factory _We are building the software factory we wanted as engineers: fast enough to keep up with ideas, careful enough to trust with real repositories._ By Reason Machines August 28, 2026 We are building the software factory we wanted as engineers: fast enough to keep up with ideas, careful enough to trust with real repositories, and opinionated enough to make the whole loop better each time it runs. Humans should still set direction, make judgment calls, and care about taste. But the mechanical middle of coding should become autonomous: reproduce, edit, verify, review, remember. That is the future of coding we are building toward. ## Code is truth In modern software development, prose and promises drift. Code and live evidence outrank everything else. Reason is architected around this principle: every proposed change must be reproduced, verified through passing tests, and accompanied by verifiable receipts before it can be merged. ## Backed by pioneers Reason is backed by Y Combinator and leading investors and operators across AI, devtools, and developer platforms. We are headquartered in San Francisco, California. --- > Source: https://reasonmachines.company/terms-of-service # Terms of Service Last updated: October 5, 2026 ## Agreement These Terms of Service ("Terms") are an agreement between you and Reason Machines, Inc. ("Reason Machines," "we," "us," or "our") governing your access to and use of Reason Machines' websites, applications, APIs, cloud sessions, local-runtime controls, and related services (collectively, the "Service"). By accessing or using the Service, you agree to these Terms. If you use the Service for an organization, you represent that you have authority to bind it, and "you" includes that organization. If you do not agree, do not use the Service. ## Eligibility You must be at least 18 years old (or the age of majority in your jurisdiction) to use the Service. By using the Service, you represent that you meet this requirement and have the legal capacity to enter into these Terms. If you are using the Service on behalf of an organization, you represent that you have authority to bind that organization. ## Accounts Some features require an account authenticated through Reason Machines' identity provider or a supported federated sign-in method. You are responsible for: - Maintaining the security of your account and credentials - All activity that occurs under your account - Notifying us immediately at contact@reasonmachines.company if you suspect unauthorized access We reserve the right to suspend or terminate accounts that violate these Terms or that we reasonably believe are being used fraudulently. ## Communications We may send service communications needed to administer your account, security, billing, and requested workflows. Where permitted by law, we may also send product updates, newsletters, tips, and feedback requests. You may unsubscribe from non-essential communications using the link in the message, but essential service communications cannot be disabled while your account is active. ## What You Can Do You may use Reason for personal or commercial projects. Features expressly identified as local operate against files on your device. Account, authentication, telemetry, model requests, connected-service actions, and any feature you direct to use Reason Cloud or a third-party provider may still transmit information as described in our Privacy Policy. Cloud sessions run in isolated compute environments managed for Reason Machines. ## Acceptable Use You agree not to: - Reverse engineer, decompile, or attempt to extract the source code of the Service - Remove or alter any branding, copyright, or proprietary notices - Use the Service for any unlawful purpose, to harm others, or to facilitate harm - Redistribute, sell, or sublicense the software or access to the Service - Use the Service to violate any third-party terms of service, including those of AI providers or connected services - Attempt to gain unauthorized access to other users' sessions, data, or accounts - Use cloud sessions for cryptocurrency mining, denial-of-service attacks, or as proxy/relay infrastructure - Exceed reasonable usage limits or abuse cloud resources in a manner that degrades the Service for others - Upload, transmit, or store malware, viruses, or other harmful code through the Service - Use the Service to send spam, phishing messages, or other unsolicited communications through connected services We may suspend or terminate access if we reasonably determine that you violated these restrictions or that suspension is necessary to prevent material harm, abuse, security risk, or legal exposure. ## Your Responsibilities - You provide your own API keys for supported AI services or use credentials managed through Reason Cloud - You comply with the terms of your AI provider and other connected services - You are solely responsible for any API costs, usage fees, or charges incurred through your use of third-party services - You keep your credentials, API keys, and account information secure - You are responsible for all content you create, send, or process through the Service, including via AI agents acting on your behalf - You comply with all applicable laws and regulations in your jurisdiction ## Connected Services You may optionally connect source-control, communication, project-management, model, MCP, or other third-party services. By connecting a service: - You authorize Reason Machines to request and store OAuth tokens for the specific permissions (scopes) you approve - You acknowledge that AI agents operating within your sessions may read, create, modify, or delete data in your connected services based on the permissions you granted and the instructions you provide - You are responsible for reviewing the permissions you grant and the actions performed by AI agents on your connected accounts - You may revoke access at any time from your Reason Machines account settings or through the connected provider Reason Machines is not responsible for any data loss, unintended modifications, sent messages, deleted files, or other consequences arising from AI agent actions on connected services, except to the extent caused by Reason Machines' breach of these Terms or where responsibility cannot be excluded by law. You should review agent actions and maintain your own backups of important data. Disconnecting a service stops future access but does not reverse completed actions or delete content already stored in Reason session history, backups, or the connected service. ## AI-Generated Output The Service facilitates interactions with third-party AI models. Regarding AI-generated content: - AI output may be inaccurate, incomplete, biased, or inappropriate. You are responsible for reviewing and validating all AI-generated content before relying on it - We do not guarantee the accuracy, quality, legality, or fitness for purpose of any AI-generated output - You are responsible for ensuring that your use of AI-generated content complies with applicable laws, including intellectual property, privacy, and professional regulations - AI agents may perform actions (send emails, modify files, execute code) based on your instructions. You accept full responsibility for these actions and their consequences ## Subscriptions and Billing Reason Cloud offers free and paid subscription plans. For paid plans: - Payments are processed securely through Stripe. We do not store your payment card details - Subscriptions renew automatically at the end of each billing period unless cancelled - You may cancel your subscription at any time through your account settings. Access continues until the end of the current billing period - Refunds are handled on a case-by-case basis at our discretion - We reserve the right to change pricing with 30 days' notice. Price changes do not affect your current billing period ## Cloud Sessions and Data When using Reason Cloud: - Sessions run in isolated containers on our infrastructure. Reason Machines personnel cannot assume your identity or enter an arbitrary workspace through a hidden product path. App-level support access requires an ordinary workspace invitation and tenant role that you control. Underlying content may still be processed or accessed as described in our Privacy Policy for infrastructure, security, legal, support, or service-quality purposes - Session workspaces are backed up to encrypted storage for continuity. We are not a backup service and do not guarantee the availability or integrity of session data - You should maintain your own backups of any important work - We may set and enforce limits on session duration, resource usage, and storage based on your subscription plan - Inactive sessions may be terminated after a reasonable period of inactivity ## Intellectual Property Your content: You retain all rights to any code, files, content, and data you create or process through the Service. We claim no ownership of your content. You grant Reason Machines and its service providers a limited, non-exclusive license to host, copy, transmit, modify, display, and otherwise process that content only as needed to provide, secure, support, and improve the Service, comply with law, and carry out your instructions. You represent that you have the rights and permissions needed to provide that content and instruct Reason Machines to process it. Our service: Reason Machines, including its software, design, branding, documentation, and infrastructure, is owned by Reason Machines and protected by intellectual property laws. These Terms grant you a limited, non-exclusive, non-transferable, revocable license to use the Service in accordance with these Terms. ## Third-Party Services The Service integrates with or relies on third-party identity, model, source-control, hosting, billing, and connected services. You acknowledge that: - Third-party services are governed by their own terms and privacy policies - We are not responsible for the availability, accuracy, or conduct of any third-party service except to the extent caused by Reason Machines' breach of these Terms or where responsibility cannot be excluded by law - Outages, changes, or discontinuation of third-party services may affect the Service, and we bear no liability for such events to the maximum extent permitted by law - Your use of third-party services through Reason Machines is at your own risk ## Service Availability We work to keep the Service running smoothly, but we do not guarantee uninterrupted or error-free operation. We may update, modify, suspend, or discontinue any part of the Service at any time, with or without notice. We are not liable for any downtime, data loss, or disruption resulting from maintenance, updates, or circumstances beyond our reasonable control. ## Disclaimer of Warranties The Service is provided "as is" and "as available" without warranty of any kind, express or implied, including but not limited to warranties of merchantability, fitness for a particular purpose, non-infringement, accuracy, or reliability. We do not warrant that the Service will meet your requirements, be uninterrupted, timely, secure, or error-free, or that any defects will be corrected. No advice or information obtained from Reason Machines shall create any warranty not expressly stated in these Terms. ## Limitation of Liability To the maximum extent permitted by applicable law, Reason Machines, its owners, developers, officers, employees, agents, and affiliates (collectively, "Reason Machines Parties") shall not be liable for any indirect, incidental, special, consequential, punitive, or exemplary damages, including but not limited to damages for: - Loss of data, revenue, profits, or business opportunities - Service interruption or downtime - Actions performed by AI agents on your behalf, including on connected third-party accounts - Unauthorized access to or alteration of your data or transmissions - Any content or conduct of any third party on or through the Service - Cost of procurement of substitute goods or services In no event shall the aggregate liability of the Reason Machines Parties for all claims arising out of or relating to these Terms or the Service exceed the greater of: (a) the total amount you paid to Reason Machines in the twelve (12) months preceding the claim, or (b) fifty U.S. dollars ($50.00). These limitations apply regardless of the legal theory (contract, tort, negligence, strict liability, or otherwise), even if Reason Machines has been advised of the possibility of such damages, and even if a limited remedy is found to have failed its essential purpose. ## Indemnification You agree to defend, indemnify, and hold harmless the Reason Machines Parties from and against any and all claims, damages, obligations, losses, liabilities, costs, and expenses (including reasonable attorney's fees) arising from: (a) your use of the Service; (b) your violation of these Terms; (c) your violation of any applicable law or regulation; (d) your violation of any third-party rights, including intellectual property or privacy rights; (e) actions performed by AI agents operating under your account or at your direction, including actions on connected services or other third-party accounts; or (f) any content you submit, post, transmit, or make available through the Service. ## Termination You may stop using the Service at any time. You may delete your account from your account settings, which will start deletion of covered data in accordance with our Privacy Policy. We may suspend or terminate access for a material breach, nonpayment, legal or security risk, abuse, fraud, or discontinuation of the Service. Where practicable, we will provide notice and an opportunity to cure. We may suspend immediately when reasonably necessary to prevent harm, comply with law, or protect the Service or others. Upon termination, your right to use the Service ceases immediately. Sections that by their nature should survive termination (including Disclaimer of Warranties, Limitation of Liability, Indemnification, and Dispute Resolution) will continue in effect. ## Dispute Resolution These Terms are governed by and construed in accordance with the laws of the State of California, United States, without regard to conflict of law principles. Before filing any formal claim, you agree to first attempt to resolve the dispute informally by contacting us at contact@reasonmachines.company. We will attempt to resolve the dispute within 30 days. If the dispute cannot be resolved informally, you agree that any legal action or proceeding shall be brought exclusively in the state or federal courts located in San Francisco, California, and you consent to the personal jurisdiction of such courts. ## Severability If any provision of these Terms is found to be unenforceable or invalid, that provision will be limited or eliminated to the minimum extent necessary so that these Terms will otherwise remain in full force and effect. ## Entire Agreement These Terms, together with the Privacy Policy, constitute the entire agreement between you and Reason Machines regarding the Service and supersede all prior agreements, representations, and understandings. ## Changes We may update these Terms from time to time. We will post revised Terms with an updated date and provide additional notice of material changes as required by law or contract. Changes apply prospectively from their stated effective date. If you do not agree to the updated Terms, you must stop using the affected Service. ## Contact Reason Machines, Inc. — [contact@reasonmachines.company](mailto:contact@reasonmachines.company) --- > Source: https://reasonmachines.company/privacy-policy # Privacy Policy Last updated: October 5, 2026 ## Scope This policy explains how Reason Machines, Inc. ("Reason Machines," "we," "us," or "our") collects, uses, shares, retains, and deletes information when you use our websites, mobile and desktop apps, APIs, cloud agents, local-runtime controls, and related services (the "Service"). ## Information We Collect Account and workspace information. We collect information such as your name, email address, profile image, account and authentication identifiers, organization memberships, roles, invitations, and settings. We receive some of this information from WorkOS or a supported federated sign-in provider. Organization administrators may also provide your email address when inviting you. Customer content. When you use cloud features, Reason Machines processes and stores the prompts, messages, agent instructions, model outputs, source code, repository contents, branches, commits, pull requests, files, attachments, screenshots, recordings, and other content needed to run and resume your sessions. This content may include personal or confidential information that you choose to provide. Connections and credentials. If you connect a source-control provider, communication or project-management service, MCP server, or another integration, we collect the connection identity, approved scopes, configuration, and encrypted access or refresh tokens. We process data returned by that service, and actions taken through it, according to the permissions and instructions you provide. API keys and other secret values are encrypted at rest. Billing and communications. We collect subscription, usage, credit, invoice, and transaction records, along with information needed to send service, security, billing, and product communications. Stripe processes payment-card and bank details; Reason Machines does not directly store full payment-card numbers. Usage, device, and diagnostic information. We collect page paths, feature usage, clicks and other interactions, account and organization identifiers, browser and device details, IP address and user agent, cookies or local-storage identifiers, request and run metadata, performance measurements, and error messages or stack traces. Operational logs may be linked to your user ID, email address, organization, repository, session, or run so we can operate and debug the Service. Voice input. If you tap the microphone control, Reason sends a short recording to its transcription service to produce an editable transcript. Reason may log limited operational metadata such as recording size, duration, and transcript length for reliability and abuse prevention. The transcript becomes customer content if you send or save it. ## Analytics and Session Replay We use cookies, local storage, and similar browser technologies for authentication, security, preferences, diagnostics, and analytics. We use PostHog for product analytics. PostHog stores a persistent analytics identifier in browser storage and a cookie and receives product-interaction events. When you are signed in, Reason Machines identifies analytics with your account ID and email address and groups activity by organization. PostHog autocapture and session replay are enabled. A replay can include page structure, navigation, clicks, content visible on screen, and text entered in ordinary non-password fields. Password fields and fields Reason Machines marks as sensitive are masked or excluded, but you should enter secret values only in fields designated for secrets. We use Axiom for application, security, performance, and error telemetry. Browser and server logs sent to Axiom can include the linked usage and diagnostic information described above. Vercel also processes website request and performance information. We do not sell this information, use it for third-party advertising, or share it with data brokers for cross-app advertising. Reason CLI and Mac app installation and update telemetry uses a random per-installation identifier. It can include the CLI or app version, release channel, operating-system family, processor architecture, installation and launch timing, launch source, update result, background-access choice, whether a retired Device installation remains, and coarse country derived at network ingress. When you sign in to the Mac app, product analytics links Mac app use to your Reason Machines account as described above. This telemetry does not include local paths, filenames, hostnames, environment names, file contents, or secret values. ## How We Use Information - Authenticate users, manage accounts, organizations, permissions, and integrations - Run, display, resume, secure, and support local and cloud agent workflows - Send code, prompts, and related context to the model and compute providers needed for a run - Connect to services and perform actions that you or your organization instruct Reason to take - Measure usage, bill accounts, provide credits, and prevent fraud or abuse - Diagnose errors, monitor reliability and security, and improve the Service, including through the limited staff review of session content described below - Communicate about the Service and comply with legal obligations ## Local and Cloud Processing Features expressly identified as local execute against files on your device. Using a Reason Machines account, the website or app, connected services, or a model provider can still transmit the account, telemetry, prompt, code, or integration data described in this policy. For cloud inference, Reason sends the context needed for your request through Concentrate to the model provider selected for the run. Depending on your selection and availability, that provider may include OpenAI or Anthropic. If you use your own provider account or credentials, the terms you accepted directly with that provider also apply. Cloud sessions run in an isolated compute environment. Reason Machines stores cloud session messages, workspaces, attachments, and checkpoints so sessions can run and resume. Customer source code and prompt context are processed by the compute and AI providers required for the run. Automated systems necessarily access this content. Reason Machines personnel cannot assume your identity or use a hidden product path to enter an arbitrary customer workspace. If app-level support requires workspace access, you must invite the designated support address through the ordinary workspace membership flow, choose its tenant role, and may remove it at any time. Separately, authorized infrastructure personnel may access underlying content only when needed to protect the Service, investigate abuse or a security incident, comply with law, or fulfill a support request, subject to access controls. A limited set of authorized Reason Machines personnel may also review session content, including prompts, messages, and the agent's tool calls and their results, to evaluate and improve the quality, reliability, and safety of the Service. These reviews use each reviewer's individual staff account, require a recorded reason, are logged, and are read-only. They exclude sessions set to Personal or Incognito and sessions that ran on your own devices, and recognizable credentials are masked. ## Connected Services Connections are optional. Reason Machines requests the OAuth permissions you approve and uses connected data to provide the features and actions you request. Depending on the connection, this can include repository data, messages, files, calendars, events, email, contacts, or other provider content. An agent can read, create, modify, send, or delete data when both the granted permission and your instruction allow it. You can disconnect a service in Reason Machines settings and can also revoke access with the provider. Disconnecting stops future access but does not automatically undo actions already completed or delete content already saved in Reason session history, backups, or the connected service. ## When We Share Information We share information only as needed for the following purposes: - Service providers: WorkOS for identity; PlanetScale for the primary database; Railway and Vercel for hosting; cloud infrastructure providers for agent compute; Tigris for object storage; PostHog and Axiom for analytics and observability; Concentrate for model routing; Deepgram for voice transcription; Stripe for billing; and Resend for transactional email. - AI and developer services: model providers such as OpenAI and Anthropic, and source-control providers such as GitHub and GitLab, when required to perform your requested workflow. - At your direction: connected communication, project-management, MCP, and other services you choose, and people or organizations with whom you share a workspace or session. - Legal, safety, and business needs: when reasonably necessary to comply with law, protect users or the Service, enforce agreements, or complete a merger, financing, acquisition, or sale with appropriate confidentiality protections. We require service providers that process personal information for Reason Machines to protect it and use it consistently with their contracts with Reason Machines and applicable law. Providers may process information in countries other than your own. ## Retention We retain information for as long as reasonably necessary to provide and secure the Service, maintain account and workspace continuity, satisfy the purpose for which it was collected, comply with legal and accounting obligations, resolve disputes, and enforce our agreements. The period depends on the information's nature, sensitivity, purpose, workspace ownership, and legal or security requirements. Credentials are retained until you remove, rotate, or revoke them or delete the owning account, connection, repository, or workspace. Certain run recordings and model request traces are scheduled for deletion after 30 days, and terminal session checkpoint bundles are scheduled for deletion after 7 days. Analytics, operational telemetry, backups, billing records, and service-provider copies follow the applicable project, provider, security, and legal retention cycles. We may retain limited records when required for tax, fraud prevention, security, dispute resolution, or another legal obligation. Content owned by a surviving workspace and actions or content already written to a connected third-party service are not deleted with an individual account. ## Security We use access controls, transport encryption, isolated compute, and encryption at rest for stored secret values. No system is completely secure, so please report suspected unauthorized access to [contact@reasonmachines.company](mailto:contact@reasonmachines.company). ## Your Choices and Rights You can update account and workspace settings, remove credentials, disconnect integrations, unsubscribe from non-essential email using the message's unsubscribe control, and revoke device or provider permissions. Depending on where you live and subject to applicable exceptions, you may also have rights to know or access, correct, export, object to or restrict processing of, or delete personal information; opt out of certain sale, sharing, targeted advertising, or profiling; appeal a denied request; and withdraw consent where processing relies on consent. We will not discriminate against you for exercising an applicable privacy right. To make a privacy request or ask us to stop non-essential analytics associated with your account, email [contact@reasonmachines.company](mailto:contact@reasonmachines.company). We may need to verify your identity. Essential account, security, billing, and operational processing may continue where necessary to provide the Service or meet legal obligations. ## Account Deletion You can request permanent account deletion from account settings. If you cannot sign in, use our public [account deletion request page](/delete-account). If you are the last owner of a workspace, you must transfer ownership or delete that workspace first. Deletion removes covered account data from Reason Machines' primary application store and starts applicable provider cleanup; it does not remove data owned by a surviving workspace or commits, pull requests, messages, files, or other actions already written to a connected third-party service. ## Children The Service is not intended for anyone under 18. If you believe a minor has provided personal information, contact [contact@reasonmachines.company](mailto:contact@reasonmachines.company). ## Changes We may update this policy as the Service or legal requirements change. We will post the revised policy with a new date and provide additional notice when a change is material. ## Contact Reason Machines, Inc. — privacy questions and requests: [contact@reasonmachines.company](mailto:contact@reasonmachines.company) --- > Source: https://reasonmachines.company/delete-account # Account Deletion Last updated: July 2026 ## Delete your Reason account You can permanently delete your Reason account and associated data from Reason on the web or in the mobile app. - [Sign in to Reason](/). - Open Dashboard for your workspace. - Scroll to Personal, then under Danger zone select Delete and complete the confirmation. This permanently deletes your Reason identity and covered data from Reason's primary systems, including sole-owned workspaces, sessions, and secrets. Reason cancels known active subscriptions for sole-owned workspaces before deleting their local data. Connected-provider apps, installations, and data may remain after deletion. Review and remove them directly with each provider. Deleted data can also remain temporarily in backups, and records required for security, fraud prevention, billing, or legal obligations are handled as described in our [Privacy Policy](/privacy-policy). ## Request deletion without signing in If you cannot access your account, email [contact@reasonmachines.company](mailto:contact@reasonmachines.company?subject=Reason%20account%20deletion%20request) from the address associated with your Reason account. Include “Reason account deletion request” in the subject. We may ask you to verify account ownership before completing the request. --- > Source: https://reasonmachines.company/zh-hans # Reason Machines 正在打造全球最出色的编程智能体。 在执行框架的新时代，快速找到运用新模型及其能力的最佳方式，就是一切。 我们的原则是让执行框架既极简又极致：极简，让我们始终行动最快；极致，让我们充分发挥前沿模型的长处。 新发布：[Reason Agent 1.0 在 DeepSWE 上的表现](/zh-hans/blog/reason-agent-deepswe) [contact@reasonmachines.company](mailto:contact@reasonmachines.company) Reason 是面向 GitHub 的自主软件工程师。连接代码库、分配问题，Reason 的云端智能体会提交附有屏幕录像、日志和可审查差异的拉取请求。您可以通过以下网页应用使用 Reason：https://reasonmachines.company/agents；面向用户的程序化访问可通过 HTTP API 实现，文档见https://reasonmachines.company/llms.txt. ## 页面 每个页面也提供简洁的 Markdown 版本：在地址后添加`.md`到任意网址后，或发送`Accept: text/markdown`。整个网站也可作为一个文件获取：https://reasonmachines.company/llms-full.txt ### 博客 - [macOS 版 Reason](https://reasonmachines.company/zh-hans/blog/ara-for-macos.md)：运行云端会话，让 Reason 使用您的本地文件、工具和端口。 - [介绍 Reason Agent 1.0](https://reasonmachines.company/zh-hans/blog/reason-agent-deepswe.md)：我们在 DeepSWE v1.1 上的首批结果。性能可与前沿执行框架竞争，相比基准测试中的对照方案，解题时间中位数最多降低 39%。 - [Rust 智能体框架之争：grok-build 与 codex](https://reasonmachines.company/zh-hans/blog/rust-harness-wars.md)：我们克隆了 xAI 的 grok-build 和 OpenAI 的 codex，逐行阅读各自的二十个子系统，探索两个实验室在构建同类编程智能体时究竟有何差异。 - [面向初创企业的 Reason](https://reasonmachines.company/zh-hans/blog/yc.md)：当前及往届 YC 公司可获得 1,000 美元额度，用于开始编程。在 Bookface 打开 Reason 优惠页面，按照兑换说明领取。 - [构建软件工厂](https://reasonmachines.company/zh-hans/blog/about.md)：我们正在打造工程师理想中的软件工厂：快到能跟上创意，稳到能处理真实代码库。 ### 法律信息 - [服务条款](https://reasonmachines.company/zh-hans/terms-of-service.md)：适用于 Reason Machines 网站、应用、API、云端会话及本地运行时控制的条款。 - [隐私政策](https://reasonmachines.company/zh-hans/privacy-policy.md)：Reason Machines 如何收集、使用、共享、保留和删除信息。 - [删除账户](https://reasonmachines.company/zh-hans/delete-account.md)：如何申请删除您的 Reason 账户及相关数据。 --- > Source: https://reasonmachines.company/zh-hans/terms-of-service # 服务条款 最后更新：2026 年 10 月 5 日 ## 协议 本服务条款（“条款”）是您与 Reason Machines, Inc.（“Reason Machines”或“我们”）之间的协议，规定您访问和使用 Reason Machines 网站、应用、API、云端会话、本地运行时控制及相关服务（统称“服务”）的权利和义务。访问或使用服务，即表示您同意本条款。如果代表某个组织使用服务，您声明自己有权约束该组织，此时“您”也包括该组织。如果不同意，请勿使用服务。 ## 使用资格 您必须年满 18 周岁，或达到所在司法管辖区的成年年龄，方可使用服务。使用服务即表示您声明自己满足此要求，并具有签订本条款的法律行为能力。如果代表某个组织使用服务，您声明自己有权约束该组织。 ## 账户 部分功能要求通过 Reason Machines 的身份提供商或受支持的联合登录方式验证账户。您需要负责： - 维护账户和凭据的安全 - 您账户下发生的所有活动 - 发现问题后立即通过以下方式通知我们：contact@reasonmachines.company如果您怀疑存在未经授权的访问 对于违反本条款，或我们合理认为存在欺诈性使用的账户，我们保留暂停或终止账户的权利。 ## 通信 我们可能发送管理您的账户、安全、计费和所请求工作流程所需的服务通知。在法律允许时，也可能发送产品更新、新闻简报、使用建议和反馈请求。您可通过消息中的链接退订非必要通信，但在账户处于有效状态期间，无法关闭必要的服务通知。 ## 您可以采取的操作 您可以将 Reason 用于个人或商业项目。明确标为本地的功能针对您设备上的文件运行。账户、身份验证、遥测、模型请求、已连接服务操作，以及您指定使用 Reason 云端或第三方提供商的功能，仍可能按隐私政策传输信息。云端会话在为 Reason Machines 管理的隔离计算环境中运行。 ## 可接受使用规范 您同意不进行以下行为： - 对服务进行逆向工程、反编译或尝试提取源代码 - 移除或更改任何品牌标识、版权或专有权声明 - 将服务用于任何非法目的、伤害他人或协助实施伤害 - 重新分发、出售或转授权软件或服务访问权 - 利用本服务违反任何第三方服务条款，包括 AI 提供商或已连接服务的条款 - 试图未经授权访问其他用户的会话、数据或账户 - 将云端会话用于加密货币挖矿、拒绝服务攻击，或作为代理及中继基础设施 - 超出合理使用限制，或以降低他人服务质量的方式滥用云资源 - 通过本服务上传、传输或存储恶意软件、病毒或其他有害代码 - 通过已连接服务发送垃圾信息、钓鱼消息或其他未经请求的通信 如果我们合理认定您违反这些限制，或为防止重大伤害、滥用、安全风险或法律风险而有必要暂停访问，我们可能暂停或终止访问。 ## 您的责任 - 您为受支持的 AI 服务提供自己的 API 密钥，或使用通过 Reason 云端管理的凭据 - 您遵守 AI 提供商及其他已连接服务的条款 - 您须独自承担因使用第三方服务而产生的任何 API 成本、使用费或其他费用 - 您妥善保管凭据、API 密钥和账户信息 - 您对通过服务创建、发送或处理的所有内容负责，包括由代表您行事的 AI 智能体处理的内容 - 您遵守所在司法管辖区的所有适用法律和法规 ## 已连接的服务 您可以选择连接源代码管理、通信、项目管理、模型、MCP 或其他第三方服务。连接服务即表示： - 您授权 Reason Machines 为您批准的特定权限范围请求并存储 OAuth 令牌 - 您确认，在会话中运行的 AI 智能体可能根据您授予的权限和提供的指令，读取、创建、修改或删除已连接服务中的数据 - 您有责任审查所授予的权限，以及 AI 智能体在已连接账户中执行的操作 - 您可随时在 Reason Machines 账户设置或已连接提供商处撤销访问权限 对于 AI 智能体在已连接服务上操作造成的数据丢失、意外修改、已发送消息、已删除文件或其他后果，Reason Machines 不承担责任，但因 Reason Machines 违反本条款而造成，或依法不得免除责任的情形除外。您应审查智能体操作，并自行备份重要数据。断开服务连接会停止后续访问，但不会撤销已完成的操作，也不会删除已存储在 Reason 会话历史、备份或已连接服务中的内容。 ## AI 生成的输出 本服务支持与第三方 AI 模型交互。对于 AI 生成的内容： - AI 输出可能不准确、不完整、存在偏见或不适当。您有责任在依赖任何 AI 生成内容之前进行审查和验证 - 我们不保证任何 AI 生成输出的准确性、质量、合法性或特定用途适用性 - 您有责任确保对 AI 生成内容的使用符合适用法律，包括知识产权、隐私和专业规范 - AI 智能体可能按照您的指令执行操作，例如发送邮件、修改文件或运行代码。您对这些操作及其后果承担全部责任 ## 订阅与计费 Reason 云端提供免费和付费订阅套餐。对于付费套餐： - 付款通过 Stripe 安全处理。我们不存储您的支付卡信息 - 除非取消，订阅将在每个计费周期结束时自动续订 - 您可随时在账户设置中取消订阅。访问权限将持续至当前计费周期结束 - 退款由我们酌情逐案处理 - 我们保留提前 30 天通知后调整价格的权利。价格变更不会影响您当前的计费周期 ## 云端会话与数据 使用 Reason 云端时： - 会话在我们的基础设施上通过隔离容器运行。Reason Machines 员工不能冒用您的身份，也不能通过隐藏的产品入口进入任意工作区。应用层面的支持访问需要由您控制的正常工作区邀请和租户角色。底层内容仍可能按照隐私政策，为基础设施、安全、法律、支持或服务质量目的而被处理或访问 - 会话工作区会备份到加密存储，以支持工作连续性。我们并非备份服务，也不保证会话数据的可用性或完整性 - 您应自行备份任何重要工作 - 我们可能根据订阅套餐，对会话时长、资源用量和存储设置并执行限制 - 会话在合理时段内持续不活跃时可能被终止 ## 知识产权 您的内容：您保留通过服务创建或处理的任何代码、文件、内容和数据的全部权利。我们不主张拥有您的内容。您授予 Reason Machines 及其服务商有限、非独占的许可，允许其仅在提供、保护、支持和改进服务、遵守法律及执行您的指令所需的范围内，托管、复制、传输、修改、展示或以其他方式处理这些内容。您声明自己具备提供这些内容及指示 Reason Machines 进行处理所需的权利和许可。 我们的服务：Reason Machines 的软件、设计、品牌、文档及基础设施均归 Reason Machines 所有，并受知识产权法律保护。本条款授予您有限的、非独占、不可转让、可撤销的许可，允许您依照本条款使用服务。 ## 第三方服务 本服务集成或依赖第三方身份、模型、源代码管理、托管、计费和其他已连接服务。您确认： - 第三方服务受其各自的条款和隐私政策约束 - 我们不对任何第三方服务的可用性、准确性或行为负责，但因 Reason Machines 违反本条款而造成，或依法不得免除责任的情形除外 - 第三方服务的中断、变更或终止可能影响本服务；在法律允许的最大范围内，我们不对此类事件承担责任 - 通过 Reason Machines 使用第三方服务的风险由您自行承担 ## 服务可用性 我们努力保持服务顺畅运行，但不保证服务不中断或无错误。我们可能随时更新、修改、暂停或停止服务的任何部分，无论是否另行通知。对于维护、更新或超出我们合理控制范围的情况所导致的停机、数据丢失或中断，我们不承担责任。 ## 免责声明 本服务按“现状”和“可用”状态提供，不附带任何明示或默示保证，包括但不限于适销性、特定用途适用性、不侵权、准确性或可靠性的保证。我们不保证服务满足您的要求，也不保证服务不中断、及时、安全或无错误，或保证任何缺陷均会得到修正。从 Reason Machines 获得的任何建议或信息，均不会构成本条款未明确规定的保证。 ## 责任限制 在适用法律允许的最大范围内，Reason Machines 及其所有者、开发者、管理人员、员工、代理和关联方（统称“Reason Machines 各方”）不对任何间接、附带、特殊、后果性、惩罚性或示范性损害承担责任，包括但不限于因以下事项导致的损害： - 数据、收入、利润或商业机会的损失 - 服务中断或停机 - AI 智能体代表您执行的操作，包括在已连接的第三方账户中执行的操作 - 未经授权访问或更改您的数据或传输内容 - 通过本服务发生的任何第三方内容或行为 - 采购替代商品或服务的费用 对于因本条款或服务引起或与之有关的所有索赔，Reason Machines 各方的累计责任在任何情况下均不超过以下两者中的较高者：(a) 提出索赔前十二（12）个月内您支付给 Reason Machines 的总金额；或 (b) 五十美元（50.00 美元）。 无论适用何种法律理论，包括合同、侵权、过失、严格责任或其他理论，这些限制均适用；即使 Reason Machines 已被告知此类损害的可能性，或有限救济被认定未能实现其根本目的，也不例外。 ## 赔偿 对于因以下事项引起的任何及所有索赔、损害、义务、损失、责任、成本和费用，包括合理的律师费，您同意为 Reason Machines 各方提供抗辩、予以赔偿并使其免受损害：(a) 您使用服务；(b) 您违反本条款；(c) 您违反任何适用法律或法规；(d) 您侵犯任何第三方权利，包括知识产权或隐私权；(e) AI 智能体在您的账户下或依您的指示执行操作，包括在已连接服务或其他第三方账户上执行的操作；或 (f) 您通过服务提交、发布、传输或提供的任何内容。 ## 终止 您可随时停止使用服务。您可在账户设置中删除账户，这会按照隐私政策启动涵盖数据的删除流程。 如发生重大违约、未付款、法律或安全风险、滥用、欺诈，或服务停止运营，我们可能暂停或终止访问。在可行情况下，我们会提供通知和补救机会。当为防止伤害、遵守法律或保护服务及他人而合理必要时，我们可能立即暂停访问。 协议终止后，您使用服务的权利立即停止。依其性质应在终止后继续有效的条款，包括免责声明、责任限制、赔偿和争议解决，将继续有效。 ## 争议解决 本条款受美国加利福尼亚州法律管辖并据其解释，不考虑法律冲突原则。 在提出任何正式索赔之前，您同意先通过以下方式联系我们，尝试非正式解决争议：contact@reasonmachines.company。我们将尝试在 30 天内解决争议。 如果无法通过非正式方式解决争议，您同意任何法律诉讼或程序均应专属提交至位于加利福尼亚州旧金山的州法院或联邦法院，并同意接受这些法院的属人管辖。 ## 可分割性 如果本条款的任何规定被认定为不可执行或无效，该规定将在最低必要范围内受到限制或被删除，而本条款的其他规定仍保持完全有效。 ## 完整协议 本条款与隐私政策共同构成您与 Reason Machines 就本服务达成的完整协议，并取代此前所有协议、陈述和谅解。 ## 变更 我们可能不时更新本条款。我们会公布修订条款并更新日期，并按法律或合同要求对重大变更作出额外通知。变更从注明的生效日期起适用于未来。如不同意更新后的条款，您必须停止使用受影响的服务。 ## 联系 Reason Machines, Inc. —[contact@reasonmachines.company](mailto:contact@reasonmachines.company) --- > Source: https://reasonmachines.company/zh-hans/privacy-policy # 隐私政策 最后更新：2026 年 10 月 5 日 ## 适用范围 本政策说明 Reason Machines, Inc.（“Reason Machines”或“我们”）在您使用我们的网站、移动和桌面应用、API、云端智能体、本地运行时控制及相关服务（统称“服务”）时，如何收集、使用、共享、保留和删除信息。 ## 我们收集的信息 账户和工作区信息。我们收集姓名、电子邮箱、头像、账户及身份验证标识、组织成员身份、角色、邀请和设置等信息。部分信息来自 WorkOS 或受支持的联合登录提供商。组织管理员在邀请您时也可能提供您的电子邮箱。 客户内容。使用云端功能时，Reason Machines 会处理并存储运行和恢复会话所需的提示、消息、智能体指令、模型输出、源代码、代码库内容、分支、提交、拉取请求、文件、附件、截图、录像及其他内容。这些内容可能包括您选择提供的个人信息或机密信息。 连接和凭据。如果您连接源代码管理提供商、通信或项目管理服务、MCP 服务器或其他集成，我们会收集连接身份、已批准的权限范围、配置以及加密的访问令牌或刷新令牌。我们依据您授予的权限和指令，处理该服务返回的数据及通过该服务执行的操作。API 密钥和其他敏感值在存储时均进行加密。 计费和通信。我们收集订阅、用量、额度、发票和交易记录，以及发送服务、安全、计费和产品通知所需的信息。支付卡和银行信息由 Stripe 处理；Reason Machines 不直接存储完整的支付卡号。 用量、设备和诊断信息。我们收集页面路径、功能使用情况、点击和其他交互、账户与组织标识、浏览器和设备详情、IP 地址与用户代理、Cookie 或本地存储标识符、请求与运行元数据、性能指标，以及错误消息或堆栈跟踪。运营日志可能与您的用户 ID、电子邮箱、组织、代码库、会话或运行关联，以便我们运营服务并排查问题。 语音输入。点击麦克风控件时，Reason 会将一段简短录音发送到转写服务，以生成可编辑的文字记录。为保障可靠性并防止滥用，Reason 可能记录录音大小、时长和转写长度等有限的运营元数据。如果您发送或保存转写文字，它将成为客户内容。 ## 分析与会话回放 我们使用 Cookie、本地存储及类似浏览器技术进行身份验证、安全保护、偏好设置、诊断和分析。我们使用 PostHog 进行产品分析。PostHog 在浏览器存储和 Cookie 中保存持久的分析标识符，并接收产品交互事件。登录后，Reason Machines 会使用您的账户 ID 和电子邮箱标识分析数据，并按组织对活动分组。PostHog 自动捕获和会话回放已启用。回放可能包含页面结构、导航、点击、屏幕可见内容，以及在普通非密码字段中输入的文字。密码字段和 Reason Machines 标记为敏感的字段会被遮蔽或排除，但您仍应仅在专门用于密钥的字段中输入敏感值。 我们使用 Axiom 处理应用、安全、性能和错误遥测。发送给 Axiom 的浏览器及服务器日志可能包含上述关联的用量和诊断信息。Vercel 也会处理网站请求和性能信息。我们不出售这些信息，不将其用于第三方广告，也不向数据经纪商共享以用于跨应用广告。 Reason CLI 和 Mac 应用的安装及更新遥测使用每次安装独有的随机标识符。遥测可能包含 CLI 或应用版本、发布渠道、操作系统系列、处理器架构、安装和启动时间、启动来源、更新结果、后台访问选择、是否仍存在已退役的 Device 安装，以及从网络入口推导出的粗略国家信息。登录 Mac 应用后，产品分析会按上述说明，将 Mac 应用的使用情况关联到您的 Reason Machines 账户。这些遥测不包含本地路径、文件名、主机名、环境名称、文件内容或密钥值。 ## 我们如何使用信息 - 验证用户身份，管理账户、组织、权限及集成 - 运行、展示、恢复、保护和支持本地及云端智能体工作流程 - 将代码、提示及相关上下文发送给本次运行所需的模型和计算服务商 - 连接服务，并执行您或您的组织指示 Reason 采取的操作 - 计量用量、向账户收费、发放额度，并防止欺诈或滥用 - 诊断错误、监测可靠性和安全性，并改进服务，包括通过下文所述的有限员工会话内容审查 - 就服务事项进行沟通，并履行法律义务 ## 本地与云端处理 明确标为本地的功能会针对您设备上的文件执行操作。使用 Reason Machines 账户、网站、应用、已连接的服务或模型提供商时，仍可能传输本政策所述的账户、遥测、提示、代码或集成数据。在云端推理中，Reason 会通过 Concentrate，将请求所需的上下文发送给本次运行选定的模型提供商。根据您的选择和可用情况，该提供商可能包括 OpenAI 或 Anthropic。如果使用您自己的提供商账户或凭据，您直接与该提供商接受的条款也同样适用。 云端会话在隔离的计算环境中运行。Reason Machines 存储云端会话消息、工作区、附件和检查点，以便会话运行和恢复。客户源代码及提示上下文由运行所需的计算和 AI 服务商处理，自动化系统也必须访问这些内容。Reason Machines 员工不能冒用您的身份，也不能通过隐藏的产品入口进入任意客户工作区。如果应用层面的支持需要访问工作区，您必须通过正常的工作区成员邀请流程邀请指定支持邮箱、选择其租户角色，并可随时移除该成员。此外，经授权的基础设施人员仅可在保护服务、调查滥用或安全事件、遵守法律或处理支持请求所必需时，按照访问控制要求访问底层内容。 少数经授权的 Reason Machines 员工还可能审查会话内容，包括提示、消息以及智能体的工具调用及其结果，以评估和改进服务的质量、可靠性和安全性。此类审查使用每位审查人员的个人员工账号，须记录审查原因，会被记录日志，并且为只读。审查不包括设置为个人（Personal）或无痕（Incognito）的会话，以及在您自己的设备上运行的会话，可识别的凭据会被遮盖。 ## 已连接的服务 连接是可选的。Reason Machines 仅请求您批准的 OAuth 权限，并使用连接的数据提供您请求的功能和操作。视连接类型而定，这些数据可能包括代码库、消息、文件、日历、事件、邮件、联系人或其他服务商内容。只有在所授予的权限和您的指令均允许时，智能体才可读取、创建、修改、发送或删除数据。 您可以在 Reason Machines 设置中断开服务连接，也可以在提供商处撤销访问权限。断开连接会停止后续访问，但不会自动撤销已完成的操作，也不会删除已保存在 Reason 会话历史、备份或已连接服务中的内容。 ## 我们何时共享信息 我们仅在以下目的所需的范围内共享信息： - 服务提供商：WorkOS 提供身份验证；PlanetScale 提供主数据库；Railway 和 Vercel 提供托管；云基础设施服务商提供智能体计算资源；Tigris 提供对象存储；PostHog 和 Axiom 提供分析与可观测性；Concentrate 提供模型路由；Deepgram 提供语音转录；Stripe 提供计费；Resend 提供事务性邮件。 - AI 和开发服务：在执行您请求的工作流程所需的范围内，使用 OpenAI、Anthropic 等模型提供商，以及 GitHub、GitLab 等源代码管理服务商。 - 依照您的指示：您选择连接的通信、项目管理、MCP 及其他服务，以及您与之共享工作区或会话的个人或组织。 - 法律、安全和业务需求：在合理必要时，为遵守法律、保护用户或服务、执行协议，或在适当保密措施下完成合并、融资、收购或出售而进行处理。 我们要求代表 Reason Machines 处理个人信息的服务商保护这些信息，并依照其与 Reason Machines 的合同及适用法律使用信息。服务商可能在您所在国家或地区之外处理信息。 ## 保留期 我们会在合理必要的期限内保留信息，以提供并保护服务、维持账户和工作区的连续性、实现收集目的、履行法律和会计义务、解决争议并执行协议。保留期限取决于信息的性质、敏感程度、用途、工作区所有权以及法律或安全要求。 凭据会保留至您将其移除、轮换、撤销，或删除其所属账户、连接、代码库或工作区。部分运行录像和模型请求追踪记录计划在 30 天后删除，终端会话检查点包计划在 7 天后删除。分析数据、运营遥测、备份、计费记录及服务商副本遵循适用的项目、服务商、安全和法律保留周期。 我们可能因税务、防欺诈、安全、争议解决或其他法律义务而保留有限记录。仍存续工作区所拥有的内容，以及已写入所连接第三方服务的操作或内容，不会随个人账户一并删除。 ## 安全 我们使用访问控制、传输加密、隔离计算，以及对存储密钥值的静态加密。没有系统是完全安全的；如怀疑发生未经授权的访问，请报告至[contact@reasonmachines.company](mailto:contact@reasonmachines.company). ## 您的选择与权利 您可以更新账户和工作区设置、移除凭据、断开集成连接、通过邮件中的退订控件退订非必要邮件，并撤销设备或提供商权限。根据您的居住地及适用例外，您也可能有权知悉或访问、更正、导出、反对或限制处理、删除个人信息；选择退出某些出售、共享、定向广告或画像活动；对被拒绝的请求提出申诉；以及在处理依赖同意时撤回同意。我们不会因您行使适用的隐私权而歧视您。 如需提出隐私请求，或要求我们停止与您账户关联的非必要分析，请发送邮件至[contact@reasonmachines.company](mailto:contact@reasonmachines.company)。我们可能需要核实您的身份。在提供服务或履行法律义务所必需的范围内，必要的账户、安全、计费和运营数据处理可能继续进行。 ## 删除账户 您可以在账户设置中申请永久删除账户。如果无法登录，请使用我们的公开[账户删除申请页面](/zh-hans/delete-account)。如果您是工作区的最后一位所有者，必须先转移所有权或删除该工作区。删除操作会从 Reason Machines 的主应用数据存储中移除涵盖的账户数据，并启动适用的服务商清理流程；但不会移除仍存续工作区拥有的数据，也不会撤回已写入所连接第三方服务的提交、拉取请求、消息、文件或其他操作。 ## 儿童 本服务不面向 18 岁以下的人士。如果您认为未成年人提供了个人信息，请联系[contact@reasonmachines.company](mailto:contact@reasonmachines.company). ## 变更 随着服务或法律要求的变化，我们可能更新本政策。我们会公布修订后的政策并注明新日期；如有重大变更，也会另行通知。 ## 联系 Reason Machines, Inc. — 隐私问题与请求：[contact@reasonmachines.company](mailto:contact@reasonmachines.company) --- > Source: https://reasonmachines.company/zh-hans/delete-account # 删除账户 最后更新：2026 年 7 月 ## 删除您的 Reason 账户 您可以在 Reason 网页端或移动应用中永久删除您的 Reason 账户及相关数据。 - [登录 Reason](/zh-hans). - 打开工作区的控制面板。 - 滚动到个人设置，在危险操作区域选择删除，并完成确认。 此操作会从 Reason 的主系统永久删除您的 Reason 身份和涵盖的数据，包括完全归您所有的工作区、会话和密钥。Reason 会先取消已知的完全归您所有的工作区的有效订阅，然后删除其本地数据。 删除后，连接服务商的应用、安装实例和数据可能仍然保留。请直接在各服务商处检查并移除。已删除的数据也可能暂时保留在备份中；出于安全、防欺诈、计费或法律义务需要保留的记录，将按照以下文件处理：[隐私政策](/zh-hans/privacy-policy). ## 无需登录即可申请删除 如果无法访问账户，请发送邮件至[contact@reasonmachines.company](mailto:contact@reasonmachines.company?subject=Reason%20account%20deletion%20request)请使用与您的 Reason 账户关联的邮箱发送，并在主题中注明“Reason account deletion request”。在完成请求前，我们可能要求您验证账户所有权。 --- > Source: https://reasonmachines.company/zh-hans/blog/ara-for-macos # macOS 版 Reason _运行云端会话，让 Reason 使用你的本地文件、工具和端口开展工作。_ 作者：Reason Machines 2026 年 8 月 29 日 macOS 版 Reason 将 Reason 会话连接到你的 Mac。经你授权，任务可以使用本地仓库、已安装的工具、文件和开发端口。 当 Reason 所需的代码或服务已在你的电脑上运行时，使用桌面应用。 ## 使用本地环境开展工作 Reason 可以读取选定的代码检出目录、运行已安装的工具并打开本地开发端口。本地访问权限按任务授予。 ![macOS 版 Reason 并排展示编程会话、本地电脑控制和代码仓库审查](/product/ara-macos-local-environment.webp) ## 审查并控制运行过程 查看当前权限、工具活动和文件变更。在同一窗口发送后续要求或停止运行。 ## 在其他设备上继续同一会话 在网页端或 iOS 上打开同一会话。消息和结果会继续保留在该会话中。 --- > Source: https://reasonmachines.company/zh-hans/blog/reason-agent-deepswe # 介绍 Reason Agent 1.0 _我们在 DeepSWE v1.1 上的首批结果已能与前沿智能体框架竞争。相较于基准对照，解题时间中位数最多降低 39%。_ 作者：Reason Machines 2026 年 9 月 14 日 [Reason Agent](/zh-hans)是 Reason Machines 打造的自主软件工程师。 今天，我们推出 Reason Agent 1.0。这是一个旨在帮助模型解决软件工程任务的编程智能体框架。我们在 DeepSWE v1.1 上的首批结果已能与前沿框架竞争，相较于 Codex 基准对照，解题时间中位数最多降低 39%。[1](#ref-1) ### DeepSWE 得分与实际运行时间中位数 DeepSWE 得分 得分更高 · 用时更少 ↗ 配置（60/71） gpt-6-astra超高高最高中低gemini-3.8-flash高中claude-opus-5最高超高高中低gpt-5.6-sol最高超高高中低claude-fable-5超高ma超高中低gpt-5.6-terra最高超高高中低glm-5.3最高kimi-k3最高grok-4.6中超高高低gpt-5.6-luna最高超高高中低gpt-5.5超高高中低gemini-3.7-flash中高低glm-5.3-flash最高deepseek-v4-pro最高claude-opus-4.8最高超高高中低qwen3.8-max超高muse-spark-1.2超高claude-sonnet-5最高超高高中低grok-4.5高deepseek-v4-flash最高muse-spark-1.1超高gpt-5.4超高gemini-3.6-flash高glm-5.2ma超高gemini-3.5-flash高kimi-k2.7-code默认claude-sonnet-4.6高gemini-3.1-pro-preview高Reason Agent Astra 超高 Reason Agent Astra · 超高 · 我们的运行结果 73.0% DeepSWE 得分 13.80 分钟 实际运行时间中位数 图 1。记录的智能体执行时间中位数，不含环境设置、排队和验证时间，并非任务的端到端延迟。公布的中位数涵盖已评分的尝试，包括失败；Reason 汇总两轮运行中记录了用时的尝试，得分分母保留全部分配任务。模型连线连接的是配置点，并非拟合出的前沿曲线。主机、服务商负载和执行条件不同，因此这些结果仅作描述，并非条件匹配的速度测试。[[1]](#ref-1) [[2]](#ref-2) ## 更早完成成功的解答 Reason 是我们的编程智能体框架，围绕一组精简的通用工具构建。模型决定如何检查仓库、修改代码并测试成果。我们让模型根据问题调整方法，而非规定一套固定的任务专用流程。 我们不认同更好的智能体必然需要更多编排这一假设。随着模型能力增强，我们认为框架应减少对工作流程的限制，更信任模型解决问题时的决策。框架的职责是提供可靠的工具和清晰的执行边界，而不是替模型决定每一步。 我们的目标是减少框架生成的上下文和强制协调调用，让模型能更直接地从调查走向经过测试的变更。这是一项设计假设，而非这些运行结果已证实的机制。工具集小，并不保证模型交互轮数更少。 ## 与已公布的参考结果并列比较 Reason 的 Astra 结果接近 Datacurve 公布的领先 Astra 配置。下表将我们的测量结果与公开参考点并列展示。[2](#ref-2)[3](#ref-3)[1](#ref-1) ### Reason Agent 能与领先框架竞争 DeepSWE v1.1 · 实测成功率 智能体 | 推理强度 | 成功率 | Reason（Astra） | 超高 | 73%67–78% | GPT-6 Astra | 超高 | 74%70–78% | Gemini 3.8 Flash | 高 | 74%70–78% | Claude Opus 5 | 最高 | 74%69–78% | GPT-5.6 Sol | 最高 | 73%68–77% | Claude Fable 5 | 超高 | 70%66–74% | 图 2。Reason（Astra）为我们的测量结果；其余行使用 mini-swe-agent 及已公布的 DeepSWE 结果。区间为 95% Wilson 置信区间，未针对重复任务的相关性进行调整。评估方法存在差异。[[1]](#ref-1)[[2]](#ref-2)[[3]](#ref-3)[[6]](#ref-6) ## 智能体框架并排比较 相同的模型和推理强度，不同的框架。在双方均成功的任务中，将 Reason 每个任务的尝试用时取平均后，其平均解题时间减少了 36.7%。Reason 总体解题时间中位数比 Codex 低约 40%（合并两轮运行后为 39%）。[1](#ref-1) 我们在相同的 113 个任务上比较 Reason 和 Codex，双方均使用 Astra 的超高推理强度。曲线合并了 Reason 的两轮运行和 Codex 的一轮运行。下方逐任务视图比较双方的第一轮运行，包括 Reason 的两次恢复，因此每个点展示双方在同一任务上的表现。 完成曲线展示了实测结果：在总体实测成功率相近的情况下，成功的尝试更早完成。解题进行到 20 分钟时，最终通过的尝试占 Reason 全部分配任务的 59.3%，Codex 为 20.4%。这些是事后统计的完成时间，并非设定了 20 分钟截止时间的实验结果。[1](#ref-1) ### Reason 与 Codex：解题时间 Astra 超高 · Reason 两轮运行，Codex 一轮运行 点击图例项目以显示或隐藏。 已确认成功率（%） 0 20 40 60 80 100 0102030405060 解题时间（分钟） 图 3。按记录的智能体用时统计最终通过的尝试，并非提前终止实验。Reason 和 Codex 的分母均保留全部分配任务。双方均使用 Astra 的超高推理强度；不同的服务商路由和执行条件限制了关于框架因果关系的推断。[[1]](#ref-1) ### Reason 与 Codex：逐任务比较 Astra 超高 · 各框架第一轮运行 · 每个点代表一个任务 左半边：Reason · 右半边：Codex abs-module-cache-flags. Reason：通过; Codex：通过 abs-stepped-slices. Reason：通过; Codex：通过 actionlint-action-pinning-lint. Reason：通过; Codex：通过 adaptix-name-mapping-aliases. Reason：通过; Codex：通过 aiomonitor-task-snapshots-diff. Reason：通过; Codex：通过 anko-default-function-arguments. Reason：通过; Codex：通过 anko-typed-variable-bindings. Reason：通过; Codex：通过 arcane-drift-detection-baselines. Reason：通过; Codex：未通过 arktype-json-schema-refs-dependencies. Reason：通过; Codex：通过 awilix-async-container-initialization. Reason：通过; Codex：未通过 bandit-incremental-cache-control. Reason：通过; Codex：通过 bandit-interprocedural-taint-checks. Reason：通过; Codex：通过 bandit-structured-nosec-directives. Reason：未通过; Codex：未通过 boa-hierarchical-evaluation-cancellation. Reason：通过; Codex：通过 cattrs-partial-structuring-recovery. Reason：通过; Codex：通过 clack-async-autocomplete-options. Reason：通过; Codex：未通过 claude-code-by-agents-recursive-delegation. Reason：通过; Codex：通过 cliffy-config-file-parsing. Reason：通过; Codex：通过 csstree-shorthand-expansion-compression. Reason：通过; Codex：通过 dasel-html-document-format. Reason：通过; Codex：通过 dateutil-rfc5545-timezone-interop. Reason：通过; Codex：通过 drizzle-orm-window-function-builders. Reason：通过; Codex：通过 dynamodb-toolbox-conditional-attribute-requirements. Reason：通过; Codex：通过 dynamodb-toolbox-lazy-recursive-schemas. Reason：通过; Codex：通过 effect-sse-httpapi-streaming. Reason：通过; Codex：未通过 eicrud-keyset-pagination-cursor. Reason：通过; Codex：通过 etree-xml-diff-patch. Reason：未通过; Codex：未通过 expr-try-catch-errors. Reason：通过; Codex：通过 fastapi-deprecation-response-headers. Reason：通过; Codex：通过 fastapi-implicit-head-options. Reason：通过; Codex：通过 fd-deterministic-multi-key-sorting. Reason：通过; Codex：通过 geo-shapeindex-serialization. Reason：通过; Codex：通过 go-critic-doc-link-checker. Reason：未通过; Codex：未通过 go-genai-streamed-function-args. Reason：通过; Codex：通过 go-git-worktree-merge-conflicts. Reason：通过; Codex：通过 goreleaser-retry-publish-auditing. Reason：通过; Codex：通过 gql-incremental-graphql-delivery. Reason：未通过; Codex：未通过 happy-dom-abort-pending-body-reads. Reason：通过; Codex：通过 happy-dom-deterministic-intersectionobserver. Reason：未通过; Codex：未通过 helm-array-merge-strategies. Reason：未通过; Codex：未通过 helm-unified-manifest-stream. Reason：通过; Codex：通过 httpx-deterministic-cookie-store. Reason：通过; Codex：通过 httpx-multipart-response-parsing. Reason：通过; Codex：通过 httpx-streaming-json-iteration. Reason：通过; Codex：通过 igel-persist-feature-schema. Reason：未通过; Codex：未通过 ink-grid-box-layout. Reason：未通过; Codex：未通过 ipython-session-bundle-replay. Reason：通过; Codex：通过 katex-multicolumn-array-spans. Reason：通过; Codex：通过 kcp-go-multiplexed-kcp-streams. Reason：通过; Codex：未通过 kea-atomic-signal-selectors. Reason：未通过; Codex：未通过 kgateway-consistent-hash-policy. Reason：通过; Codex：通过 kombu-single-active-consumer-priority. Reason：通过; Codex：通过 kombu-virtual-queue-dead-lettering. Reason：通过; Codex：通过 koota-composite-trait-aspects. Reason：通过; Codex：通过 koota-deferred-mutation-buffer. Reason：未通过; Codex：通过 koota-entity-snapshot-rollback. Reason：通过; Codex：未通过 koota-pair-relation-tracking. Reason：未通过; Codex：未通过 koota-query-predicates. Reason：通过; Codex：未通过 kysely-window-grouping-helpers. Reason：通过; Codex：通过 langchain-request-coalescing. Reason：通过; Codex：通过 mashumaro-flattened-dataclass-fields. Reason：通过; Codex：通过 meriyah-explicit-resource-declarations. Reason：未通过; Codex：未通过 mnamer-daemon-watch-lifecycle. Reason：通过; Codex：通过 mobly-grouped-test-barriers. Reason：通过; Codex：通过 narwhals-rolling-window-suite. Reason：通过; Codex：通过 numba-stencil-boundary-modes. Reason：通过; Codex：通过 obsidian-linter-auto-table-of-contents. Reason：未通过; Codex：未通过 obsidian-linter-link-format-conversion. Reason：未通过; Codex：未通过 obsidian-linter-scoped-ignore-markers. Reason：通过; Codex：通过 ofetch-per-origin-circuit-breaker. Reason：通过; Codex：通过 onedump-dump-encryption-pipeline. Reason：未通过; Codex：未通过 opa-rego-rule-profiling. Reason：未通过; Codex：未通过 opa-template-string-reconstruction. Reason：通过; Codex：未通过 optique-conditional-option-dependencies. Reason：未通过; Codex：通过 oxvg-structural-selector-preservation. Reason：未通过; Codex：通过 participle-grammar-conflict-analysis. Reason：未通过; Codex：未通过 pebble-durability-wait-apis. Reason：通过; Codex：通过 pest-character-class-coalescing. Reason：未通过; Codex：未通过 prometheus-transactional-reload-status. Reason：未通过; Codex：通过 prometheus-typed-label-sorting. Reason：通过; Codex：通过 psd-tools-blend-range-api. Reason：通过; Codex：通过 pwntools-tube-multiplexing. Reason：通过; Codex：未通过 python-statemachine-state-data-scoping. Reason：未通过; Codex：未通过 query-persist-restored-query-state. Reason：通过; Codex：通过 quill-shared-toolbar-focus. Reason：通过; Codex：通过 returns-validated-error-accumulation. Reason：通过; Codex：通过 scc-bounded-memory-spilling. Reason：通过; Codex：通过 scriggo-method-declarations. Reason：通过; Codex：通过 skrub-duration-encoding. Reason：通过; Codex：通过 sql-formatter-bigquery-pipe-formatting. Reason：通过; Codex：通过 sqlfmt-create-table-ddl-formatting. Reason：未通过; Codex：未通过 sqlite-utils-safe-import-checkpoints. Reason：通过; Codex：通过 superjson-error-stack-serialization. Reason：通过; Codex：通过 task-task-graph-export. Reason：通过; Codex：通过 tengo-callable-instance-isolation. Reason：通过; Codex：通过 tengo-destructuring-bindings. Reason：通过; Codex：通过 termenv-preserve-ansi-resets. Reason：未通过; Codex：未通过 testem-bail-on-test-failure. Reason：未通过; Codex：未通过 testem-per-launcher-reports. Reason：未通过; Codex：通过 textual-kitty-key-phases. Reason：通过; Codex：通过 textual-richlog-follow-state. Reason：通过; Codex：通过 tomlkit-toml-table-converters. Reason：通过; Codex：通过 true-myth-iterable-collection-combinators. Reason：通过; Codex：通过 ts-pattern-match-each. Reason：未通过; Codex：未通过 updo-policy-alerting. Reason：通过; Codex：通过 valibot-recursive-schema-composition. Reason：通过; Codex：通过 vitest-duration-sharding. Reason：通过; Codex：通过 vulture-persistent-analysis-cache. Reason：未通过; Codex：未通过 wasmi-trap-coredumps. Reason：通过; Codex：通过 wazero-multi-module-snapshots. Reason：未通过; Codex：未通过 yaegi-go-embed-directives. Reason：通过; Codex：通过 yjs-map-conflict-detection. Reason：通过; Codex：通过 ytt-jsonpath-query-api. Reason：通过; Codex：通过 双方通过 Reason 通过 Codex 通过 双方均未通过 图 4。比较 Reason 第一轮运行与 Codex，包括 Kea 环境设置失败后的重试，以及 Pwntools 未改动补丁的重新评分。蓝色表示通过。悬停或聚焦数据点，可查看双方结果。[[1]](#ref-1) ## 下一步 这些结果仍属初步结论。我们认为，编程智能体应围绕精简的通用工具集构建，让能力强的模型有更大自由度解决任务。接下来，我们将在匹配条件下使用新任务检验这一原则，逐项改变工具和编排方式，观察更少的上下文和调用能否在不牺牲正确性的前提下加快结果产出。 ## 评估说明 - 这些测量评估的是提取出的编程框架，而非托管产品。不同的服务商路由和执行基础设施限制了因果推断及严格等同条件的声明。[1](#ref-1) - 解题时间不含环境设置、试验前排队和验证。图 4 仅展示 Reason 第一轮运行；汇总图表保留两轮运行。[1](#ref-1) - 9 月 14 日更新：Reason 在 226 个分配任务中通过 165 次、失败 61 次，没有未评分结果。Kea 在环境设置失败后的新尝试中失败；Pwntools 的未改动补丁重新评分后通过。没有替换任何已评分的失败结果。JSON 保留了原始结果和恢复详情。公布的得分沿用 Datacurve 的评分规则。[1](#ref-1)[2](#ref-2)[3](#ref-3) - 运行结果、汇总得分和解题时间见结果 JSON。[1](#ref-1) ## 参考资料 \[1\] Reason Machines。DeepSWE v1.1。[运行结果与汇总数据（JSON）](/research/reason-agent-deepswe/deepswe.json). \[2\] Datacurve。DeepSWE v1.1。[报告方法](https://deepswe.datacurve.ai/blog/deepswe-v1-1)·[已公布排行榜](https://deepswe.datacurve.ai/artifacts/v1.1/leaderboard-live.json). \[3\] Datacurve。DeepSWE v1.1 运行过程数据。[试验浏览器](https://deepswe.datacurve.ai/data/v1.1)·[原始试验索引](https://deepswe.datacurve.ai/artifacts/v1.1/trials.json). \[4\] Datacurve。DeepSWE：衡量前沿编程智能体。[评估方法](https://deepswe.datacurve.ai/blog/deepswe#evaluation-harness). \[5\] Datacurve / Harbor。DeepSWE v1.1，113 个任务。[数据集](https://hub.harborframework.com/datasets/datacurve/deep-swe-1-1/latest). \[6\] NIST/SEMATECH。《统计方法电子手册》，第 7.2.4.1 节。[比例的 Wilson 置信区间](https://www.itl.nist.gov/div898/handbook/prc/section2/prc241.htm). --- > Source: https://reasonmachines.company/zh-hans/blog/rust-harness-wars # Rust 智能体框架之争：grok-build 与 codex _我们克隆了 xAI 的 grok-build 和 OpenAI 的 codex，逐行阅读双方各二十个子系统，探究两个实验室在构建同类编程智能体时究竟有何不同。_ 作者：Reason Machines 2026 年 7 月 23 日 世界上能力最强的终端编程智能体中，有两个都用 Rust 编写，以单一二进制文件发布，并在相隔几个月内开源：xAI 的[grok-build](https://github.com/xai-org/grok-build)，即 grok CLI，以及 OpenAI 的[codex](https://github.com/openai/codex)。我们克隆了两者，安排约一百个分析智能体逐行阅读二十个子系统，只为回答一个问题：两个实验室构建同一种工具时，设计取向究竟在哪里分道扬镳？ 一个令人不太自在的起点，就在 grok-build 自己的[THIRD-PARTY-NOTICES](https://github.com/xai-org/grok-build/blob/main/THIRD-PARTY-NOTICES)中。它移植了以下项目的工具实现：[openai/codex](https://github.com/openai/codex)和[sst/opencode](https://github.com/sst/opencode)。这并非完全独立开发的竞争，而是后继者对前辈的重新组合，因此两者的差异反而更值得研究。 ## 相同骨架，不同取向 并排阅读会发现，两者的骨架几乎相同。一个会话 actor 位于类型化消息边界之后。同一个二进制文件提供多种轻量入口：交互式终端界面、无界面执行模式以及编辑器服务器。核心循环也一样：调用模型采样，运行它请求的工具，重复直到停止。双方都将只追加日志持久化为可靠的事实来源，都在上下文窗口填满时压缩对话，都支持恢复会话和创建分支。 因此，有意思的不是架构形状，而是构建出这套骨架之后，各团队在哪些方面坚持不妥协。Codex 做减法，grok-build 向外扩展。 ## Codex 做减法 Codex 出自一个拥有自家模型、追求控制力、安全性和精简功能面的团队。它的本能是减少选项。它移除了旧的聊天通信协议，只保留一条通往模型的路径。它用单一的推理强度选项替代温度和 top-p 采样。它不打算成为服务商市场，只内置少数第一方后端。 这种专注换来了关键环节的深度：从源码构建、保证确定性的三平台沙箱；按轮次启用功能门控工具规划的策略引擎；包围模型每一条命令的审批与提权层；以及由审查强制执行、禁止改写模型可见历史的规则。阅读 Codex，就像在观察一个为大量从未见过的机器构建平台的团队。 ## grok-build 向外扩展 grok-build 在架构上更有野心，有些地方也更不成熟。它没有做减法，而是走得更远。机器级主进程通过[智能体客户端协议](https://agentclientprotocol.com)服务多个热客户端。这一新兴跨厂商标准在这里是内部通信协议，而非外加适配。与服务商无关的采样器分发到三种不同的模型协议，新增服务商靠配置而非代码。记忆是一个真正的子系统：将 Markdown 文件建立索引用于检索，而非对历史轮次简单执行 grep。它还内置了[tree-sitter](https://tree-sitter.github.io/tree-sitter/)代码图用于结构化导航，并提供能跨运行积累的会话记忆。 这套拓扑面向的是运行在他人环境中、生命周期超越任何单一模型的工具。grok-build 防范厂商锁定，正如 codex 防范不可信机器。 ## 评分表 在我们比较的二十个子系统中，codex 在十一个方面的设计更强，八个方面各有千秋，一个方面打成平手。grok-build 没有在任何维度上明确胜出，原因也很实在：codex 是成熟的上游产品，grok-build 则是 v0.1 alpha 版本，源码清楚地反映了这一点。 但那八个难分高下的方面，恰恰体现了野心的回报。更广的模型支持、记忆检索、编辑器协议、云端执行和精确费用统计，都因为尝试得更远，给出了真正不同、有时更好的答案。纸面上的十一比零，实际上讲述的是成熟度与探索野心之间的竞赛。 11 项 Codex 占优 8 项 grok-build 各有千秋 1 项完全持平 子系统 | 结论 | 架构分析 | 沙箱与隔离 | Codex | Codex 提供从源码构建的三操作系统沙箱（macOS/Linux/Windows）。grok-build 覆盖两种操作系统。 | 安全与策略门控 | Codex | Codex 强制执行逐轮功能门控与审批升级提示。grok-build 的安全钩子失败时默认放行。 | 模型协议多路复用 | 各有千秋 | grok-build 清晰地在 OpenAI、Anthropic 和 xAI 之间分发。Codex 限定使用第一方通信协议。 | 智能体客户端协议（ACP） | 各有千秋 | grok-build 采用新兴标准 ACP，实现跨厂商的客户端—守护进程架构。 | 记忆与状态积累 | 各有千秋 | grok-build 为 Markdown 文件建立检索索引。Codex 则重放并压缩原始对话记录。 | 代码图与 Tree-Sitter | 各有千秋 | grok-build 包含原生 Tree-sitter AST 导航 crate，用于构建仓库结构图。 | 技能互操作性 | 各有千秋 | grok-build 可直接使用 Claude 和 Cursor 的 SKILL.md 文件夹，无需手动转换。 | 技能深度与能力 | Codex | Codex 技能是完整的可执行包，包含单元测试脚本和界面元数据。 | 工具治理中间件 | Codex | Codex 使用严格的平台策略防护与验证中间件包围执行过程。 | 对话记录完整性 | Codex | Codex 强制保持模型历史不可变、不可改写，以支持可复现的审计。 | 配置加固 | Codex | Codex 内置加密签名密钥。grok-build 的签名机制在 v0.1 中并未实际生效。 | 无界面执行引擎 | Codex | Codex 的无界面 CI 执行器具有确定性，并经过实战验证。 | 费用与 Token 统计 | 各有千秋 | grok-build 在混合后端间维护精确的逐轮 Token 与金额账本。 | 编辑器服务器集成 | 各有千秋 | grok-build 原生集成为 ACP 语言服务器。 | 错误恢复与继续执行 | Codex | Codex 具有完善的循环检测和自动熔断机制。 | 终端界面（TUI）安全性 | Codex | Codex TUI 完全使用安全 Rust。grok-build 使用原始指针转换。 | 发布打包与可移植性 | 持平 | 两者都编译为单一、静态、独立的 Rust 二进制文件，无需外部运行时。 | 提示裁剪与压缩 | Codex | Codex 的积极上下文压缩保留了推理轨迹。 | 云端执行通信 | 各有千秋 | grok-build 的守护进程拓扑支持向远程无界面工作进程交接任务。 | 遥测与可观测性 | Codex | Codex 的结构化事件追踪支持端到端回放验证。 | ## 技能、记忆与工具 在技能方面，两者采用相同的[SKILL.md](https://github.com/anthropics/anthropic-quickstarts/blob/main/computer-use-demo/skills.md)约定，并逐步披露信息：在真正打开技能前，上下文中只有名称和描述。grok-build 优先考虑互操作性，可原样读取 Claude 和 Cursor 的技能目录，让已有技能库直接可用。codex 则追求深度，将技能视为包含脚本、参考资料和界面元数据的文件夹，是更丰富的能力单元。 在记忆方面，codex 保留原始对话记录，并据此重建状态。grok-build 使用专用 crate，持久化可由人编辑的 Markdown 并建立检索索引。在工具方面，codex 将工具拆分为模型可见的规范和执行运行时，再叠加完整的治理中间件。grok-build 将工具建模为严格的三 crate 契约，职责分离清晰，但安全层更轻。相同模式反复出现：codex 构建更深入的系统，grok-build 构建更便携的系统。 ## Alpha 版本的代价 grok-build 的扩展野心也有代价，阅读源码就能看清。配置签名子系统带有两千行测试，却没有内置公钥，因此实际上未生效。钩子失败时默认放行，意味着安全钩子崩溃会悄悄停止执行约束。它的沙箱覆盖两个平台，而 codex 覆盖三个。终端界面中还有不安全的指针转换，靠人工约束而非类型系统保障安全。 对 alpha 版本而言，这些都并非致命问题。多数只是年轻项目的常见特征：威胁模型思考走在了已发布的安全加固之前。但这也是先追求更大架构野心、再补足成熟度的代价。 未生效的配置签名：安全缺口 拥有 2,000 行测试夹具，但发布的二进制中没有任何内置根公钥。签名约束形同虚设。 失败时放行的安全钩子：策略风险 如果外部安全前置钩子发生 panic 或崩溃，执行会继续而不经检查，而不是终止当前轮次。 两操作系统与三操作系统沙箱：覆盖缺口 仅覆盖 macOS 和 Linux。Windows 开发环境中无沙箱执行（Codex 提供 AppContainer 隔离）。 不安全的 TUI 指针转换：Rust 不变式 终端事件处理循环中手动转换指针，依赖开发者自律，而非借用检查器。 ## 我们为什么关心 Reason 是一家围绕智能体框架构建的公司，因此深入阅读两个世界顶尖框架，并非出于无所事事的好奇心。我们不断重新认识到同一条经验：模型是你租用的商品，框架是你拥有的产品。Codex 和 grok-build 是对框架形态的两种可信选择：一个做减法、重安全，一个做加法、重可移植性。两者之间的差距，正是整个领域的关键。 ## 这项分析如何完成 我们克隆了两个仓库，直接阅读源码，而非依赖文档或记忆。多智能体流程分派了约一百个分析智能体：架构绘图者、在二十个维度上分别阅读两仓库的四十个子系统阅读者、二十个一对一比较者，以及综合分析小组；每个智能体返回的数据都经过模式验证。成熟度和边界情况的判断基于所阅读的代码，而非厂商声明。grok-build 版本为 v0.1.220 alpha，codex 使用当时的 main 分支。 --- > Source: https://reasonmachines.company/zh-hans/blog/yc # 面向初创企业的 Reason _在读及往届 YC 公司可领取价值 1,000 美元的额度，用于开始编程。在 Bookface 打开 Reason 优惠页面，按照兑换说明领取。_ 作者：Reason Machines 2026 年 9 月 3 日 参加 YC 是为了与用户交流、寻找产品市场契合点并快速推进，而不是将整天耗费在机械性的错误修复、样板工作和过时依赖上。在读及往届 YC 创始人可直接通过 Bookface 兑换 Reason 专属额度，加快工程迭代。 Reason 作为团队中的自主软件工程师，处理重复性的编程工作，并提交经过验证的拉取请求供你审查。 ## 专心经营公司，把代码交给 Reason 无论是复现偶发的结账错误、追踪身份验证 webhook，还是发布集成端点，Reason 都会在隔离的云环境中自主执行任务，返回测试通过、可供审查的拉取请求。 ![Reason 正在执行软件任务并审查代码变更](/product/reason-work.webp) ## 关注真正重要的指标 专注于决定公司发展的指标——月度经常性收入、用户留存、激活率和客户对话——让 Reason 同时维护代码库。 ![Reason 模型用量与活动概览](/product/reason-security.webp) ## 如何兑换 在读 YC 创始人和往届成员可前往[Bookface 上的 Reason 优惠页面](https://bookface.ycombinator.com/deals/6551)，或使用公司邮箱登录以下地址领取优惠：[reasonmachines.company/login](/login). --- > Source: https://reasonmachines.company/zh-hans/blog/about # 构建软件工厂 _我们正在构建工程师心中理想的软件工厂：快到能跟上想法，也足够严谨，值得托付真实的代码仓库。_ 作者：Reason Machines 2026 年 8 月 28 日 我们正在构建工程师心中理想的软件工厂：快到能跟上想法，严谨到值得托付真实仓库，并且有足够明确的设计原则，让整个流程每运行一次都变得更好。 人类仍应设定方向、作出判断并把握品味。但编程中间那些机械性的工作应该自主完成：复现、修改、验证、审查、记忆。这就是我们正在构建的编程未来。 ## 代码才是事实 在现代软件开发中，文字和承诺会逐渐失真。代码和实时证据高于一切。Reason 围绕这一原则构建：每项提议的变更都必须复现问题、通过测试验证，并附带可核实的证据，才能合并。 ## 先行者的支持 Reason 得到了 Y Combinator 以及人工智能、开发工具和开发者平台领域领先投资者与运营者的支持。总部位于美国加利福尼亚州旧金山。