1.4 KiB
Agent Types & Roles (Practical Taxonomy)
Use this skill to choose the right kind of agent workflow for the job.
Common agent “types” (in practice)
1) Chat assistant (no tools)
Best for: explanations, brainstorming, small edits. Risk: can hallucinate; no grounding in repo state.
2) Tool-using single agent
Best for: well-scoped tasks where the agent can read/write files and run commands. Key control: strict DoD gates + minimal permissions.
3) Planner + Executor (2-role pattern)
Best for: medium complexity work (multi-file changes, feature work). Flow: Planner writes plan + acceptance criteria → Executor implements → Reviewer checks.
4) Multi-agent (specialists)
Best for: bigger features with separable workstreams (UI, backend, docs, tests). Rule: isolate context per role; use separate branches/worktrees.
5) Supervisor / orchestrator
Best for: long-running workflows with checkpoints (pipelines, report generation, PAD docs). Rule: supervisor delegates, enforces gates, and composes final output.
Decision rules (fast)
- If you can describe it in ≤ 5 steps → single tool-using agent.
- If you need tradeoffs/design → Planner + Executor.
- If UI + backend + docs/tests all move → multi-agent specialists.
- If it’s a pipeline that runs repeatedly → orchestrator.
Guardrails (always)
- DoD is the truth gate.
- Separate branches/worktrees for parallel work.
- Log decisions + commands in AGENT_LOG.md.