Agent

An agent is a named, role-scoped expert prompt that the orchestrator can hand a task to during a loop. Agents come in two flavors:

• Universal agents — test-strategist, security-privacy, and others that ship with every project.
• Domain agents — generated by bootstrap for the specific domains of your codebase (payments-domain, auth-service, analytics-pipeline, and so on).

Each agent has a stable role, a human-readable name, a description, and a versioned prompt. The desktop client reads agents from .claude/agents/<name>.md at execution time; the agents API stores the same roster at the organization level so it can be versioned, reused across repos, and pulled into context packs for judges.

Every prompt or name change creates a new version. Older versions remain queryable so loops in flight can pin the prompt they started with.

See Agents API and the bootstrap plugin.