Cursor Releases Agent SDK: Agents as Code, Deployed With a Single Command

Cursor has opened up Agent SDK to everyone. It is the internal framework the company uses to build its own managed agents: Bugbot, Security Reviewer and Rollouts. The idea is simple. You describe an agent in code and keep it in a repository, where it goes through review like any other software. It reaches production with the agent-sdk deploy command. The SDK is already generally available.

Until now, teams had two options. They could either take an off-the-shelf product built around someone else's workflow, or write an agent from scratch. Cursor is offering a third: take the same scaffolding its own engineers use and adapt it to your own rules.

An agent as a small software project

The structure of an agent resembles an ordinary repository. Instructions are written in Markdown, tools are described as TypeScript functions, and the connections to external services sit alongside the evals used to test the agent. Each file is responsible for something the agent knows or can do.

The main practical consequence is that changes to an agent's behavior go through the same pull request, the same commit history and the same review as the rest of the code. The team can read and edit the policies the agent follows, rather than guessing what happens inside someone else's black box.

The basic agent definition takes a few lines. In the example from the documentation, the code calls defineAgent from the @cursor/agent-sdk package and sets the name Robo, the model composer-2.5 with the parameters effort: high and fast: true, and a local working directory. From there, you can iterate in a local playground, run evals and only then deploy.

Reacting to events and working as a team of agents

Agents connect to the places where the team works and trigger on events. In the example from the announcement, an agent listens for pull request events on GitHub, GitLab or Bitbucket, assesses the risk of the change, routes it to the right team and approves on its own the changes that fit within policy.

Inside Cursor, one team had built up a large, fast-growing knowledge base of lessons from past decisions and wanted every PR checked against it. Using Agent SDK, the team built kb-reviewer in about a day. The agent runs on every PR in the main codebase. It flags changes that contradict a recorded lesson, links to the note behind each comment and nudges engineers to resolve the discrepancies.

For complex tasks, an agent can hand work off to another specialized agent with its own tools and sessions, or split a task among subagents that work in parallel. Evals and event replay are built in, so critical scenarios can be tested before rollout.

Deployment and first customers

Run from the project directory, the agent-sdk deploy command builds the agent, assigns it a URL and runs it on Cursor's infrastructure. Secrets and integrations are configured in the dashboard, and logs, evals and the playground work the same way as they do locally. The alternative is to run the agent yourself. Each agent is a single Node process, and you can package it in a container.

The first notable customer named is PGA Tour. According to Michael Gyarmathy, principal engineer at PGA America, the team built every component of its software development factory on Cursor's infrastructure, and the built-in evals let it quickly test new ideas and reliably optimize existing capabilities.

What's in the release

• A framework for defining agents in code: prompts in Markdown, tools in TypeScript
• Starter templates and skills that walk a coding agent through the development process
• Triggers for GitHub, GitLab and Bitbucket events
• Handoff of tasks between specialized agents, and parallel subagents
• Built-in evals and event replay
• A local playground and deployment to Cursor's infrastructure with a single agent-sdk deploy command
• A dashboard for secrets, integrations and logs
• Self-hosting: a single Node process or a container

Cursor also suggests you skip writing the agent by hand: you can simply point your own coding agent at the documentation, and it will help you build the bot together with you.

For teams that already automate review, compliance and incident analysis, the practical takeaway is this: start with one repetitive process that has clear rules, such as checking PRs against an internal knowledge base, as kb-reviewer does. Set up evals before the first deployment, keep the agent in your main repository and decide in advance whether you need Cursor's hosting or your own container. The agent then becomes as manageable a part of your system as the rest of your code.