Cursor’s Agent SDK: Build an Agent as a Folder of Code, Eval It, Deploy It

Cursor’s Oct 20 changelog introduced the Agent SDK. An agent is a project folder: instructions.md is the prompt, each file in tools/ is a typed action, skills/ holds on-demand procedures, mcp-connections/ pulls in MCP tools, and channels/ defines where it listens.

What you get

• Start with npx @cursor/agent-sdk init ./my-bot, then agent-sdk dev for a local playground. Templates include a security reviewer, a PR autofixer and issue triagers.
• agent-sdk deploy runs it on Cursor, or you can host it yourself.
• Agents can wake on GitHub, GitLab and Bitbucket events, webhooks or a cron schedule.
• Evals send a fixed message and check which tools ran and what came back.

Review before you install: read the docs and any script or package yourself before running it, and try it somewhere you can afford to break first.

My take

Agents as code you can test, with evals built in, is the right direction.

God Bless, and Use the Source!

References & Further Reading

• Cursor changelog: Agent SDK