Introducing the Agent SDK
Cursor’s Agent SDK lets teams define, evaluate, and deploy event-driven agents as code, with local evals, event replay, subagents, and an option to run on their own infrastructure.
Cursor’s **Agent SDK** is generally available for building agents in code and deploying them to Cursor infrastructure with `agent-sdk deploy`. An agent project includes Markdown instructions, TypeScript tools, external connections, and evals; it can also run as a Node process in a container.
Builders can use it to automate team workflows such as PR review, then version and review the agent like other software. The SDK includes evals and event replay, and agents can delegate work to specialized agents or parallel subagents.
Cursor’s **Agent SDK** is generally available for building agents in code and deploying them to Cursor infrastructure with `agent-sdk deploy`. An agent project includes Markdown instructions, TypeScript tools, external connections, and evals; it can also run as a Node process in a container. Builders can use it to automate team workflows such as PR review, then version and review the agent like other software. The SDK includes evals and event replay, and agents can delegate work to specialized agents or parallel subagents. The source describes one internal knowledge-base reviewer built in about a day, but gives no general build-time estimate. Deployment to Cursor is one option; self-hosting is also supported, with infrastructure and operational details left to the docs.
Cursor’s SDK packages agent instructions, tools, connections, evals, replay, and deployment or self-hosting options into a versionable project. It adds a concrete build-and-operate path for team workflows, while the single internal example does not establish general implementation time or comparative performance.