Cursor Turns Agents into Code, With Built-in Evals

Cursor has officially released the Agent SDK. Cursor makes an AI coding editor, and this time it's packaging agent capabilities into an SDK for developers.

Building an agent used to feel like setting up a manual workshop: prompts scattered across docs, tool functions copied all over the place, and nobody really knew how the whole thing worked. Now Cursor defines the agent as code. Prompts are written in Markdown, tools in TypeScript, and deployment can run on their hosted infrastructure or on your own servers.

What does this mean for developers? The most immediate change: agent configs can go into a git repo. Code review, version rollback, branch management — agents get the same treatment as ordinary code. When an agent is code, you can diff it, and you can discuss its behavior changes in a PR.

Another key piece is built-in evals. Every change can be run against an eval set, and the results are quantified directly. Tuning agents used to rely on gut feel — it seemed smarter, or maybe it didn't. Now there are numbers. This makes agent iteration feel more like engineering and less like black magic.

The official video shows an Agents as code repository, including directories such as review-agent, code-review-agent, and compliance-team-PR. Each agent corresponds to a specific type of task, and managing agents looks just like managing service code.

For people who don't write code, this news may not land with much impact. But the trend worth noting is that agents are shifting from chat toys into testable, deployable engineering systems. That inflection point is more worth watching than benchmark scores climbing the leaderboards.