Agent overview

Asterweave ships thirteen specialist agents — Markdown files under plugins/asterweave/agents/ with a frontmatter tool allowlist and a scoped system prompt. Each graph node delegates to one, either a plugin agent by default or a project-specific agent configured through agent routing.

Executor / evaluator separation

Read-only analyzers and reviewers cannot edit files; only implementers and testers can write, and only within their assigned scope. This separation means the agent that reviews a change is never the agent that wrote it.

Agent	Read	Write	Role
repo-analyzer	✅	—	Repository discovery
requirements-challenger	✅	—	Requirements challenge
architect	✅	—	Implementation planning
implementer	✅	✅	Bounded implementation
test-engineer	✅	✅ (tests only)	Test design and execution
verification-engineer	✅	—	Runtime acceptance verification
staff-reviewer	✅	—	Staff-level code review
security-reviewer	✅	—	Independent security review
pr-engineer	✅	— (Git/MCP only)	PR, pipeline, and work-item operations
failure-analyst	✅	—	Diagnosing repeated failures
scaffold-architect	✅	—	Repository scaffold design
scaffold-auditor	✅	—	Independent scaffold audit
orchestrator	✅	✅	Coordinates every other agent

Reviewer → implementer loop

A finding never gets fixed by the agent that raised it — it routes back through the orchestrator to implementer, then the change is re-verified and re-reviewed.

Repository-specific agents

A repository can define its own agents under .claude/agents/*.md and route a graph stage to one through .claude/asterweave.json. Create one only for genuinely bounded, recurring domain expertise a generic Asterweave agent cannot reasonably provide — for example finance-domain-reviewer for regulated settlement logic. Do not create one for generic code review, generic security review, or generic implementation; Asterweave already provides those. See Repository-specific agents and Asterweave vs. the repository.

Related

Agent routing — how a graph stage resolves to a plugin or project agent.