uygnoey Instructions
You should help the user decide whether a multi-agent system is warranted, and if so, propose a practical architecture.
- Start by attempting a single-agent solution.
- Only recommend multi-agent if at least one strong driver applies:
- Context protection (avoid context pollution)
- Parallelization (explore in parallel; thoroughness over speed)
- Specialization (focused toolsets, prompts, or deep domain expertise)
- Use context-centric decomposition:
- Prefer boundaries that keep related context together (e.g., feature + tests).
- Avoid boundaries that require heavy shared state or sequential handoffs.
- If adopting multi-agent, propose:
- Orchestrator responsibilities
- Subagent responsibilities and context boundaries
- What minimal artifacts flow between them (summaries, diffs, checklists)
- Add a verification subagent when independent validation is valuable.
- Require explicit acceptance criteria.
- Guard against “early victory” by forcing comprehensive checks (including negative tests).
Refer to the companion reference for decision criteria and the companion examples for implementation sketches.
Examples
- Decision & architecture proposal: see references/decision-framework.md
- Context protection pattern: see examples/context-protection.md
- Parallel research pattern: see examples/parallel-research.md
- Specialization routing pattern: see examples/specialization-routing.md
- Verification subagent loop: see examples/verification-subagent.md