Cursor Agent SDK Launches

• On 20 October 2026 Cursor made the Agent SDK available: an agent is a folder of code, tested with evals and deployed to Cursor or your own infrastructure^[1]
• Start with npx @cursor/agent-sdk init ./my-bot, run agent-sdk dev for a local playground, and optionally pass --template for a security reviewer, owners-policy approver, PR autofixer or Linear or Jira triager^[1]
• Cursor's blog calls it the framework it uses for its own managed agents Bugbot, Security Reviewer and Rollouts^[3]

Notable features

• Project layout: instructions.md is the prompt, each file in tools/ is a typed action, skills/ holds on-demand procedures, mcp-connections/ adds MCP tools, channels/ defines where the bot listens^[1]
• Channels: GitHub, GitLab and Bitbucket PR, comment and CI events, Slack mentions and DMs, webhooks and cron; each conversation keeps its own resumable session^[1]
• Evals assert which tools ran and the reply shape; agent-sdk github replay turns a real PR into signed webhook payloads for testing^[1]
• Not in the changelog: a Cursor team built kb-reviewer in about a day to check every PR against a lessons-learned base, and PGA America said evals let it iterate fast on its software factory^[3]
• RuntimeWire says teams remain responsible for the policies and evaluations that make agents dependable^[2]