Multi Agent Orchestration

Multi-Agent Orchestration

Every handoff is a context loss. Justify it.

The default assumption is that more agents means more capability. The actual outcome is more tokens, more latency, more variance, and harder debugging — unless the handoff buys something a single agent with tools cannot.

When to use this skill

• The user is proposing a multi-agent system (Crew, Swarm, Supervisor, A2A).
• The user has a multi-agent system and reports loops, drift, runaway cost, or "no one ever decides".
• The user is comparing "subagents" vs "tools" for the same workload.
• The user is using a framework's multi-agent abstraction without a reason that survives "why not a tool?".

The handoff test

Before splitting work into a second agent, the handoff must clear all three:

1. Different context. The second agent needs context the first agent must not see (token budget, role boundary, privacy).
2. Different tools or model. The second agent uses tools, a model tier, or constraints the first cannot.
3. Clear termination. The second agent returns a single typed result; control returns to a known place.

If any of the three fails, you want a tool call, not an agent.

Decision flow

1. Can a single agent with the right tools do the job? → yes? Stop. Add the tool.
2. Do you need to enforce a role boundary the prompt can't enforce? (e.g. critic must not see actor's chain-of-thought) → maybe a second agent. Or use structured prompting.
3. Do sub-tasks need genuinely different models / token budgets / safety policies? → multi-agent justified.
4. Are agents collaborating or delegating? → delegation only. Free-form collaboration loops are an anti-pattern.
5. Who owns termination? → exactly one coordinator. Peer-to-peer termination is a footgun.

Patterns that work

• Supervisor + workers. Supervisor owns state, dispatches a worker per turn, workers are stateless. Returns to supervisor with a typed result.
• Pipeline (planner → executor → verifier). Strict ordering, explicit handoffs, no back-edges except reject → re-plan.
• Triage → specialist. First agent classifies; second agent does the work. Specialist never re-classifies.
• Critic-actor (Reflexion). Same task, separate roles, hard iteration cap (≤3).

Patterns to refuse

• Peer-to-peer chat. Agents discussing the task to "reach consensus" burn tokens and rarely converge.
• Round-robin debate. Same as above with a turn schedule.
• Hierarchical-of-hierarchies (team-of-teams). Debuggability collapses. Use as last resort, never as default.
• Agent-as-tool with implicit shared memory. Side-channel state ruins replayability.

Handoff design rules

• Typed payload. Handoff message is a typed schema, not free text.
• One return path. The receiver returns to a known parent or terminates the run.
• Termination contract. Max turns, max wall-clock, max cost — all three, all enforced.
• No back-channels. Agents communicate only via the documented handoff API.
• Trace the handoff. Every handoff is a span. See [[agent-observability]].

Anti-patterns to flag immediately

• Multi-agent for parallelism. If sub-agents don't need different context/tools/model, you want concurrent tool calls.
• Agent personas as design. "Senior engineer reviews junior engineer's code" — these are roles, not architectures. Use a single agent with a critic step.
• Free-text handoff. "Manager: do this thing — Worker: ok" with no schema. Loses fidelity at every step.
• No termination. Agents that can hand off without bound. Always set step + cost ceilings.
• Cross-agent shared mutable memory. Context poisoning at scale.
• Framework-driven design. "We're using CrewAI, so we have a crew." Pick the pattern first, the framework second.

Questions to ask the user

1. What does the handoff buy — different context, different tools, different model, different policy?
2. What is the typed payload at each handoff?
3. Who owns termination and what are the limits?
4. How does the system fail gracefully when an agent loops or hangs?
5. What does a single trace look like across agents — can you replay it?
6. If you collapsed this to one agent with tools, what would break?

If the last answer is "nothing would break", collapse it.

The hard line

Every additional agent must clear the handoff test or be deleted. Multi-agent is a debt instrument; interest is paid in tokens and latency on every call.

Why this exists

The multi-agent boom has produced a generation of systems that are slower, dumber, and more expensive than the single-agent versions they replaced. The systems that work treat handoff as a structural decision — typed payload, single termination, hard limits — not a vibes-level "let agents collaborate".

References

• references/patterns.md — supervisor, pipeline, triage, critic-actor — control-flow sketch and failure modes for each.
• references/handoff-payloads.md — typed schemas for common handoffs (triage → specialist, planner → executor, actor → critic).
• references/frameworks.md — CrewAI, AutoGen, OpenAI Swarm, LangGraph multi-agent — where each helps, where each leaks.

Related skills

• Multi-agent is one shape of [[agent-architecture-patterns]] — start there.
• Each agent's tools are still [[tool-use-schema-design]].
• Cost scales by agent-call, not by agent-count — [[agent-cost-modeling]].
• Latency budget must include handoff round-trips — [[latency-budgeting]].
• Handoffs must be traced — [[agent-observability]].
• Loops and hangs are guardrail concerns — [[guardrails-and-safety]].