-
Notifications
You must be signed in to change notification settings - Fork 0
Agents
7 specialized subagents with orchestration chains.
| Agent | Role | Memory | Model |
|---|---|---|---|
| researcher | Read-only codebase exploration | Transactional | haiku |
| architect | Design decisions, tradeoff analysis | Persistent | opus |
| implementer | Code + tests | Persistent | sonnet |
| code-reviewer | Review by severity (critical/warning/suggestion) | Persistent | sonnet |
| security-auditor | Vulnerability scanning (OWASP top 10) | Persistent | opus |
| test-runner | Run tests + report coverage | Transactional | sonnet |
| session-reviewer | Post-session analysis | Transactional | sonnet |
researcher → architect → implementer → test-runner → code-reviewer
Defined in .claude/rules/agents.md. Claude follows the delegation decision tree:
- Single-file fix / quick question → handle directly
- Research-heavy → delegate to researcher
- Code changes + tests → delegate to implementer
- Security concern → delegate to security-auditor
- Multi-component refactor (>3 files) → evaluate Agent Teams
- Code review → delegate to code-reviewer
- Architecture decision → delegate to architect
Agents with persistent memory write learnings to .claude/agent-memory/<agent-name>/. These survive across sessions and prevent repeated mistakes.
Transactional agents (researcher, test-runner, session-reviewer) execute and report without accumulating knowledge.
All agents include a 5K token output budget constraint. Summaries over raw output — agents don't echo full files or dump raw grep results.
Agent definitions support these frontmatter fields:
| Field | Description |
|---|---|
allowed-tools |
Tool whitelist (use allowed-tools:, not tools:) |
model |
Model override: haiku, sonnet, opus
|
effort |
Reasoning budget: normal, high, max (architect=high, security-auditor=max) |
context |
fork — skill executes in isolated subagent (used by 5 heavy skills) |
color |
UI color hint |
Model assignments follow the criteria in template/rules/model-routing.md (installed to every bootstrapped project). See Model Routing for the full selection criteria and escalation rules.
For large tasks touching 3+ independent components, multiple agents can work in parallel using isolation: "worktree". A lead agent coordinates; teammates each own a distinct file set.
To continue a subagent's work, use SendMessage({to: agentId}) — never spawn a new agent for follow-ups.