Agents

Agents are the individual workers that make up your Subterranean team. Each agent can be configured with a default model (Claude Opus, Claude Sonnet, or GPT) and a tool permission preset — Orchestrate, Build, Design, or Operations — that tailors its access to match the role it plays in your project.

Beyond manual runs, agents can wake themselves up through automation. A job runs an agent on an hourly, daily, or weekly schedule to proactively check project status, while event triggers start a run when rows change in one of your tables or when webhook events arrive from an authorized GitHub repository. Full configuration guidance is coming soon.

When a preset is too coarse — because the tool in question is run_bash or http_request and what matters is the arguments, not the tool — an agent can carry a policy document instead: versioned rules with a reason attached to each decision. See agent permissions by job for complete, tested examples.