Cursor Launches Agent SDK to Let Engineering Teams Define, Evaluate, and Deploy AI Agents as Versioned Markdown and TypeScript Code
The Agent SDK is now available. Define agents as code, with prompts in Markdown and tools in TypeScript. Deploy them on our hosted infrastructure, or run them on your own. Built-in evals let you measure every change and make your agents better over time.
Context
Exposing the internal harness behind Bugbot and Security Reviewer lets engineering organizations version control agent prompts, tools, and evaluation suites alongside application code rather than managing brittle dashboard configurations.
Cursor previously built internal managed agents such as Bugbot, Security Reviewer, Approval Bot, and Rollouts on a proprietary orchestration harness to automate pull request reviews and release workflows. Until now, teams wanting custom workflow agents had to either configure simple prompt triggers in a web dashboard or build custom agent loops from scratch.
Deep dive
Cursor announced general availability of the Agent SDK on October 26, opening up the internal software framework that powers its own managed agents including Bugbot, Security Reviewer, Approval Bot, and Rollouts. The toolkit lets developers structure an AI agent like a standard software repository where system instructions live in Markdown files and custom tools, MCP connections, hooks, channels, and subagents are written in TypeScript.
Because every prompt, tool definition, and permission boundary is stored as code, engineering teams can run agent changes through standard pull request reviews and version control. The SDK includes built in evaluation suites and event replay out of the box, enabling developers to test how an agent responds to GitHub, GitLab, Bitbucket, or Slack events locally before shipping updates to production.
Once validated in the local playground, teams can deploy an agent to Cursor's managed cloud infrastructure with a single command or package the agent as a standalone Node process inside a container to run on self hosted servers. Cursor noted that early enterprise adopters such as the PGA Tour are already using the Agent SDK and its evaluation harness to automate software development workflows.
What to expect next
Treating agent system prompts, TypeScript tool definitions, and evaluation benchmarks as versioned repository artifacts bridges the gap between brittle no code chat bots and bespoke infrastructure engineering. As engineering organizations deploy specialized autonomous agents for pull request review, incident response, and security compliance, integrated evaluation replay and one command cloud or container hosting will rapidly become standard across developer platforms.