Multi-Agent Orchestration
Overview
The skill governs the supervisor side of multi-agent work: deciding when a task warrants splitting across multiple agents, designing non-overlapping roles and scopes, budgeting context and constraints between models, writing delegation prompts with robust checkpoint contracts, verifying agent deliverables independently, recovering failed agents without restarting from scratch, and keeping a ledger of operational costs and leverage. The reader is an agent or supervisor orchestrating a team of agents on a complex task.
When to use
- A task is too large or complex for a single agent context and requires parallel exploration or disjoint specialization.
- Producer-validator separation is required so that the agent generating an artifact does not grade its own work.
- Multiple independent sub-problems must be investigated concurrently to meet tight wall-clock timeframes.
- A multi-step autonomous workflow requires robust failure recovery, checkpointing, and cost tracking.
When NOT to use
Managing handoff artifacts between people and sessions is outside; that is handled by ac-handoff. Triaging one human's inbound request queue is outside; that is handled by ac-inbound-triage. Prompting a single model or tuning instruction lines for an individual agent is outside; that is handled by ac-prompt-engineer.
Subskill index
| Leaf | Domain segment | Scope (one line) |
|---|---|---|
subskills/when-to-split.md |
collaboration/when-to-split |
Deciding when a multi-agent split is justified by independent contexts, parallel wall-clock need, or producer-validator separation |
subskills/role-and-scope.md |
collaboration/role-and-scope |
Designing the split: one outcome per agent; non-overlapping scopes; producer-is-not-validator invariant; least-privilege assignment |
subskills/context-budget.md |
collaboration/context-budget |
Minimal sufficient context, not the transcript; constraints passed verbatim; direction-dependent handoffs; file-based state |
subskills/delegation-prompts.md |
collaboration/delegation-prompts |
Worker prompt structure: role preamble, mission in one line, verbatim constraints, output contract, file checkpointing, no questions |
subskills/verification-loops.md |
collaboration/verification-loops |
Accepting agent work: verify file on disk; validator differs from producer; adversarial passes for load-bearing claims; cost trade-off |
subskills/failure-and-recovery.md |
collaboration/failure-and-recovery |
Failure classes: transient vs latent/context faults; resumption from partial files with continuation brief; escalation paths |
subskills/ledger-and-leverage.md |
collaboration/ledger-and-leverage |
Running the loop: per-task cost and retry records; irreversible acts stay human-gated; sequential heavy jobs; leverage ratio measurement |
Status and honest limits
- Seedling: assembled from public research canon on agent workflows and internal operational playbooks.
- Multi-agent orchestration incurs inherent coordination overhead; unwarranted splits degrade performance rather than improving it.
- Supervisor judgment remains critical; rules and decision trees narrow choices but do not replace operational oversight.
- Cost and latency models are estimates based on standard multi-model handoffs and vary by provider infrastructure.
⚠️ Open question: no standardized benchmark universally quantifies the exact coordination threshold where multi-agent splits outperform single-agent execution across all domains.