agent-coordination

Files (1)

• SKILL.md

SKILL.md (first lines)

# Agent Coordination

Use this only in the parent coordinator. The `delegate` tool comes from the separate `pi-delegate` package and owns execution mechanics; this skill owns model-routing judgment. If the tool is missing, report that and let the calling skill finish its own cleanup; don't substitute another mechanism.

## Load Routing Policy

Before selecting a delegated model or reasoning level, read `~/.pi/agent/agent-routing.md` once for the current coordination task.

- An explicit user model or reasoning choice wins.
- If the file is absent or unusable, omit `model` and `reasoning` so `delegate` inherits the parent route.

## Shape the Work

Before choosing a route, decide whether the work has independent partitions whose results are cheaper to combine than asking one agent to inspect everything. Lookup, extraction, source review, summarization, distillation, bounded analysis, and independent review lenses often split well. Sequential decisions, shared evolving state, and tightly coupled implementation usually do not.

Fan out to the natural partitions. For independently inspectable sources or artifacts with compact results, prefer one task per item. Group items only when they share context or per-result overhead would make one task per item wasteful. Do not impose an arbitrary smaller batch or create duplicate work to fill the available concurrency. Put partitions with the same model, reasoning, skills, tools, context, cwd, and workspace mode into as few `delegate({ tasks: [...] })` calls as the current batch limit permits. Use separate delegate calls only when that limit requires it or the partitions' resources differ.

Give each agent a bounded question and ask for only the evidence or findings needed for synthesis. The parent combines the results first. Escalate only when synthesis itself requires stronger judgment or leaves a cross-cutting question that does. Do not rerun every partition.

## Select the Route

1. Identify the delegated task's quality requirements, context demands, latency needs, and how readily its result can be checked. Judge the task itself, not just its parent workflow.
2. Apply the configured model preferences and choose a supported reasoning level. If the policy links model evidence, read it when its conditions apply; examples inform the choice rather than restrict it to listed pairs.
3. Pass the selected `model` and `reasoning` explicitly to `delegate`. If the selected model is unavailable, surface the error rather than silently substituting another model.
4. If agents require different routes or resources, use separate `delegate` calls because batch resources are shared.

Follow the `delegate` and `delegate_control` tool guidance for lifecycle, waiting, and delivery. Do not duplicate status requests or poll for completion.

From ryanmazzolini/pi-skills, MIT.

More from ryanmazzolini/pi-skills