Agents defined as a "folder"
Cursor, the U.S. maker of an AI coding tool, announced on X on Oct. 26 (Oct. 27 Japan time) the general availability of "Agent SDK," a development kit for building custom AI agents. Agents are defined as code, run through evaluation tests, and run on Cursor's platform or in users' own environments.
Each agent is contained in a single folder. Instructions are written in a Markdown document, and each tool the agent uses is a separate TypeScript file. Instruction guides loaded when needed and MCP settings for connecting to external tools are placed in the same folder. As with ordinary code, changes can be compared, tested and version-controlled.
Cursor says it also builds its own code review bot "Bugbot" and agents for tasks such as security checks with the same development kit.
Triggered by pull requests and Slack
Agents are triggered by pull requests and comments on GitHub, GitLab and Bitbucket, test results, mentions in Slack, notifications from external services, and set times. Templates are provided for roles such as a security reviewer, an automatic pull request fixer and an issue triager.
Evaluation tests send fixed requests and check which tools the agent used with what inputs and whether the response format is correct. Real pull requests can also be replayed to serve as test material.
Model fees go through Cursor
Agents can run either on Cursor's platform, where they can be deployed with a single command, or on users' own servers. In both cases, however, model processing goes through Cursor, and even when using Anthropic's Claude or OpenAI's GPT, the bill comes from Cursor. Third-party models are charged at API list prices on plans of $30 a month and above, and Team and Enterprise add a further $0.25 per million tokens.
Anthropic also offers its own development kit for its models, "Claude Agent SDK." With that kit, usage fees are paid to Anthropic or to the clouds of Amazon or Google, and only Claude can be used.
