Cursor Agent SDK Opens: Agents Move Into the Code Repository, and Prompts Also Need to Go Through PRs

Cursor's Agent SDK is now open. The official positioning is straightforward: Agent as code. Instructions are written in Markdown, tools in TypeScript, and triggers, evaluations, deployment, and version management all follow software projects. Developers do not need to build an agent runtime from scratch; once defined, they hand it to Cursor to execute.

Even without writing code, you can think of it as writing a robot that can do work as a project folder. Inside are instructions, tools, tests, and deployment configs. In the official demo, an agent is just a few folders: review-agent, code-review-agent, compliance-team-PR, on-call-alert-trigger.

What it can do now includes: code review, security checks, issue triage, and other automated workflows; being triggered from GitHub, GitLab, Bitbucket, Slack, webhooks, or scheduled tasks; connecting to external tools through MCP, skills, and subagents; using built-in eval to test behavior and replay real PR events; deploying to Cursor-hosted infrastructure or running on your own servers; and putting agents, tools, instructions, and tests under version control. This SDK is built on the same agent infrastructure Cursor uses to run Bugbot, Security Review, Approval Bot, and Rollouts.

The interesting part is not that there is one more agent platform. Cursor pulls agents back from prompt experiments in a console into the code repository. In the past, many teams' agents were a black box in a dashboard: changing a prompt was based on feel, and when something went wrong, people relied on chat logs. Now instructions and tools can go through PRs, be reviewed, and be rolled back. Nevena Cankovic said in a reply that treating agents like ordinary software, versioned, testable, and reviewable, is a bigger upgrade than black-box prompts in a dashboard. Tinker also mentioned that Markdown instructions can be reviewed in PRs like code, which by itself can fix a lot of prompt drift.

But do not overlook the toolchain choice. Vineer noticed that Cursor uses TypeScript for tools, not Python. The choice is not surprising: Cursor's developer ecosystem leans more toward frontend and full-stack. For data and ML teams, Python toolchains are more familiar, so migration costs should be calculated before adopting it. The SDK lowers the cost of building a runtime, not every team's cost of reworking their tools.

What is really worth watching is eval and replay. When an agent changes one instruction, whether the effect gets better or worse cannot rely on feeling that it seems smarter. Built-in evaluation and replaying real PR events give each change a comparison. Ofek Shaked asked a key question: how much do replayed PR events differ from real traffic? That question will determine the credibility of eval. Another developer named intelligence put it more directly: before writing agent instructions, get types and tests right first. He mentioned that after Cursor agent fixed a Sentry filter, to make 3 tests actually fail, it forced the helper to return true. This detail is more useful than writing a few more Markdown rules.

Deployment options are also worth mentioning. Cursor provides hosted infrastructure and also allows running on your own servers. NZ followed up in a reply asking whether it can run on more than just Cursor. For teams with high compliance and security requirements, self-hosting is a prerequisite for whether it can be used at all. Raymond Thurman said this is very similar to Vercel's Eve agents: Markdown instructions, tools as code, and no need to build your own runtime. The category is converging in the same direction; Cursor's difference is that it has already grown alongside workflows like code review, alerts, and PRs.

The point of Agent SDK is to make agents manageable, testable, and rollback-able software assets. Agents will not suddenly become smarter because of the SDK. For teams, the first step is to pick a high-frequency scenario with clear boundaries, such as code review or issue triage, and first get instructions, tools, tests, and the trigger chain working end to end. Automating every process at once just moves prompt roulette from the dashboard into the repository.

Docs: https://cursor.com/docs/agent-sdk