Agent Team

Run team-shaped work across sub-agents and deliver one synthesized result that states what was covered, what each material claim rests on, whether it was independently verified, and what remains a gap. Delegation is assumed authorized.

Orchestration mechanics — packet wording, parallel dispatch, pacing, how many agents run at once — are the model's own judgment. This skill fixes the bar the result must meet: honest coverage, independent verification of what the answer relies on, candidate-shaped decisions, and isolated mutators.

Delegation

Delegate genuinely independent, sizeable tracks: wide investigations, noisy exploration where only a compact result matters, or work that benefits from an independent perspective. Keep local anything finishable in a handful of tool calls and any blocking decision the main agent must make now. When one sub-agent can cover a set of named roots, use one.

Before launch, state a short plan: the units, the named roots each covers, who verifies which claims, and the stop rule. Every packet carries the shared context agents would otherwise rediscover — systemic risk, invariants, known not-a-bug items — and asks for evidence with source pointers. Sub-agent reports are evidence, not truth.

Hard Gate: Coverage

Activation: before launch, for any run whose answer speaks about a surface — audits, inventories, maps, sweeps, research.

Required: