Agents as code
Write the prompt, model, permissions, and budget in a file next to your code, then push to deploy. Every change goes through a pull request, so the team can review how each agent is configured.
An agent is a YAML file in your repo: the prompt, the model, the sandbox, the permissions, the budget. Like infrastructure-as-code, for your agents.
$
Define the agent in YAML, live on merge
The version of the file on your default branch is the live agent. Merge a change and the next session uses it. There is no deploy step and nothing to restart.
The pull request that deploys an agent
Add your own MCP servers, scope what the agent can do on GitHub, and set spend limits per session, day, week, and month. It all sits next to the prompt, so a reviewer sees what an agent can reach and spend before it runs.
session: permissions: github: repositories: [api-repo] permissions: contents: read pull_requests: writeAccess and spend declared beside the prompt