Agent Routing¶
- Handle clear, small tasks directly. When routing helps, reuse available metadata and inspect only what is missing. Load selected skill bodies.
- Use the narrowest relevant expertise: second-brain for PARA, product-manager for product uncertainty and domain experts for implementation. Skills can guide the current session without switching agents.
- Check actual host capabilities and instructions. Delegate only when tools exist, authorization permits it and the benefit exceeds coordination cost. Otherwise apply specialist guidance in the current session.
- Supply a bounded goal, context pointers, acceptance criteria, allowed files and expected evidence. Avoid concurrent writers to the same file.
- Inspect returned changes and checks, resolve contradictions and update an existing work record when present. A successful tool call does not prove completion.
Selection examples¶
| Outcome | Agent | Initial skills |
|---|---|---|
| Add a Python API endpoint | python-expert | python-patterns, api-design |
| Improve an interface | frontend-expert | frontend-patterns |
| Clarify a product increment | product-manager | product-discovery |
| File notes in a private vault | second-brain | obsidian-para |
| Coordinate several deliverables | joerg | work-management, agent-routing |
| Review a risky change | code-reviewer | code-review |
Host boundaries¶
- Exported profiles do not themselves create sessions. OpenCode's available native task/subagent tools determine delegation behavior.
- Do not promise that the user can type into a child session in every OpenCode Web release. Session visibility and interaction depend on the host UI.
- Do not invent tool names, models, account access or background execution. Select models by available capabilities and workload, not a hardcoded obsolete price table.
- Honor host permissions. A read-only reviewer must not become a writer merely because its Markdown describes a repair.
Canonical source: skills/agent-routing/SKILL.md.