[{"data":1,"prerenderedAt":56},["Reactive",2],{"news-story-news-20261026-cursor-agent-sdk-agents-as-code-hosted-evals":3},{"id":4,"title_zh":5,"title_en":6,"summary_zh":7,"summary_en":8,"takeaways_zh":9,"takeaways_en":15,"content_zh":21,"content_en":22,"external_links":23,"source":34,"category":41,"impact_score":42,"views_count":43,"likes_count":44,"related_entities":45,"tags":50,"image_url":55},"news-20261026-cursor-agent-sdk-agents-as-code-hosted-evals","Cursor 正式推出 Agent SDK：用 Markdown 写提示、TypeScript 写工具，把 Agent 当代码管理，可托管部署也可自建，内置评测","Cursor launches Agent SDK: agents as code with Markdown prompts, TypeScript tools, built-in evals, and hosted or self-run deployment","Cursor 10 月 26 日宣布 Agent SDK 正式可用：Agent 的系统提示写成 Markdown，工具用 TypeScript 定义，可以部署在 Cursor 的托管基础设施上，也可以在自己的机器上运行，并内置评测来衡量每次改动。官方文档称它使用的是驱动 Bugbot、Security Review、Approval Bot 和 Rollouts 的同一套 Cursor SDK 运行框架，可以由代码仓库事件、Issue 跟踪器、Slack 或定时任务触发，部署后在 Deployed Agents 中管理。官方未公布定价和跑分。","On Oct 26 Cursor announced that its Agent SDK is generally available. An agent's system prompt is written in Markdown and its tools in TypeScript; agents can be deployed on Cursor's hosted infrastructure or run on your own machines, and built-in evals measure every change. Cursor's docs say it uses the same Cursor SDK harness that powers Bugbot, Security Review, Approval Bot and Rollouts, can be triggered from source-control events, issue trackers, Slack or schedules, and is managed under Deployed Agents after deployment. Cursor has not published pricing or benchmarks.",[10,11,12,13,14],"同源框架：官方文档称 Agent SDK 使用驱动 Bugbot、Security Review、Approval Bot、Rollouts 这 4 个 Cursor 自家 Agent 的同一套运行框架","项目结构：bot/instructions.md 是必需的常驻系统提示，bot/tools/*.ts 每个文件一个带类型的工具，另有 skills、mcp-connections、subagents、channels、hooks、schedules 等目录，评测放在 evals/*.eval.ts","触发与部署：可由代码仓库、Issue 跟踪器、Slack、Webhook 或定时任务触发；可选 Cursor 托管或自建运行，部署后在 Deployed Agents 管理","命令行：agent-sdk validate、run、eval 分别用于校验、实跑一轮和跑评测，官方称只有 run 和 eval 需要 Cursor 凭证","官方未公布定价和跑分",[16,17,18,19,20],"Same harness: Cursor's docs say Agent SDK runs on the harness behind four of its own agents: Bugbot, Security Review, Approval Bot and Rollouts","Project layout: bot/instructions.md is the required always-on system prompt, each bot/tools/*.ts file is one typed tool, with folders for skills, MCP connections, subagents, channels, hooks and schedules, and evals live in evals/*.eval.ts","Triggers and deployment: runs can be triggered from source control, issue trackers, Slack, webhooks or schedules; deploy on Cursor-managed hosting or self-host, then manage under Deployed Agents","CLI: agent-sdk validate, run and eval check the project, run a real turn and run evals; Cursor says only run and eval need a Cursor credential","Cursor has not published pricing or benchmarks","### 核心背景与行业痛点\n团队自己做 Agent 时，提示词散落在后台配置里，工具接入各写各的，改一句提示就可能悄悄把效果改坏，又没有回归测试兜底。Cursor 此前已经在内部用同一套框架做了 Bugbot、Security Review 等 Agent，10 月 13 日 AICoder 也报道过其 BDK 的扩展。这次 Cursor 在 X 上宣布 Agent SDK 正式可用，把这套做法开放给所有开发者：Agent 的行为写成仓库里的文件，像普通代码一样评审、测试和发版。\n\n### 架构亮点与底层机制\n按官方文档，一个 Agent SDK 项目是一个由 Markdown 和 TypeScript 组成的文件夹：bot/instructions.md 是必需的常驻系统提示，bot/tools/ 下每个 .ts 文件定义一个带类型的工具，另有 skills（按需加载的流程）、mcp-connections（接 MCP 服务）、subagents（子 Agent）、channels（HTTP 入口）、hooks（只读事件订阅）和 schedules（定时运行）等目录。开发者写提示和工具，模型循环、工具调用、会话、沙箱、密钥管理和部署由 Cursor 负责；可选 Cursor 托管，也可以在自己的基础设施上跑。触发源包括代码仓库事件、Issue 跟踪器、Slack 线程、自定义 Webhook 和定时任务。\n\n### 权威 Benchmark 与实测跑分对比\n这是开发工具发布，官方未公布任何跑分，也未在我们核实的文档中给出定价。所谓「内置评测」指的是项目自带的 evals/*.eval.ts 回归测试，用来给自己的 Agent 打分，不是对外可比的基准成绩。\n\n### 开发者实战落地与开箱指南\n官方文档推荐的流程是：先用 agent-sdk validate 校验项目，用 agent-sdk tool call 直接调用工具确认返回的证据正确，再用 agent-sdk run 实跑一轮、查看打印出的会话轨迹，最后 agent-sdk eval 跑冒烟评测，并补一遍 TypeScript 类型检查。官方称只有 run 和 eval 需要 Cursor 凭证。也可以让仓库里的编程 Agent 按 bdk-create-agent 等技能脚手架出第一版。适合 PR 审查、发布审批、值班助手这类需要接 GitHub、Slack 和 Issue 系统的内部 Agent。","### Background and pain points\nWhen teams build their own agents, prompts often live in admin consoles, each tool integration is hand-rolled, and a one-line prompt edit can quietly break behavior with no regression tests to catch it. Cursor already used one harness internally for agents like Bugbot and Security Review, and AICoder covered an extension for its BDK on Oct 13. Now Cursor has announced on X that Agent SDK is generally available, opening that approach to all developers: agent behavior becomes files in the repo that are reviewed, tested and shipped like ordinary code.\n\n### Architecture highlights\nPer the official docs, an Agent SDK project is a folder of Markdown and TypeScript. bot/instructions.md is the required always-on system prompt, each .ts file under bot/tools/ defines one typed tool, and there are folders for skills (procedures loaded on demand), MCP connections, subagents, channels (HTTP routes), hooks (observe-only event subscribers) and schedules. Developers write the prompts and tools while Cursor handles the model loop, tool calls, sessions, sandboxing, secrets and deployment, on Cursor-managed hosting or your own infrastructure. Triggers include source-control events, issue trackers, Slack threads, custom webhooks and schedules.\n\n### Benchmarks\nThis is a developer-tool release. Cursor published no benchmarks, and the docs we checked list no pricing. The built-in evals are each project's own evals/*.eval.ts regression tests for scoring your agent, not comparable public benchmark results.\n\n### Getting started for developers\nThe docs suggest running agent-sdk validate on the project, calling tools directly with agent-sdk tool call to confirm they return the right evidence, running one real turn with agent-sdk run and reading the printed session trace, then running smoke evals with agent-sdk eval, followed by a TypeScript check. Cursor says only run and eval need a Cursor credential. You can also have a coding agent in your checkout scaffold a first version with skills such as bdk-create-agent. It fits internal agents for PR review, release approval or on-call help that connect to GitHub, Slack and issue trackers.",[24,28,31],{"title":25,"url":26,"type":27},"Agent SDK docs (Cursor)","https://cursor.com/docs/agent-sdk","docs",{"title":29,"url":30,"type":27},"Agent SDK project layout (Cursor docs)","https://cursor.com/docs/agent-sdk/reference/project-layout",{"title":32,"url":33,"type":27},"Build with a coding agent (Cursor docs)","https://cursor.com/docs/agent-sdk/building-with-agents",{"platform":35,"author_name":36,"author_handle":37,"author_avatar":38,"tweet_url":39,"published_at":40},"x","Cursor","@cursor_ai","https://unavatar.io/x/cursor_ai","https://x.com/cursor_ai/status/2114793714531102901","2026-10-26T18:58:07.000Z","tooling",85,125,0,{"models":46,"agents":47,"vendors":49},[],[36,48],"Bugbot",[36],[36,51,52,53,54],"Agent SDK","agents as code","evals","TypeScript","https://pbs.twimg.com/amplify_video_thumb/2114793685091270656/img/GNmvZQ3NLo9g5Lp6.jpg",1793718863888]