multi-agent-orchestration

Multi-Agent Orchestration

Purpose: Coordinate multiple specialized agents to solve tasks no single agent can handle reliably. Scope: Topologies, handoff protocols, framework selection, failure modes, anti-patterns.

---

When to Use This Skill

• Task spans multiple specialties (research + code + review) and a single agent loop loses focus
• Need explicit role separation for auditability or compliance
• Long-horizon tasks where one agent's context budget is insufficient
• Cross-organization agent communication (A2A protocol)

When NOT to Use Multi-Agent

• Single-domain task -- a well-prompted single agent is cheaper and more reliable
• Latency-sensitive (<1s) -- handoffs add round-trips
• Tasks solvable by tool calls alone -- prefer tool-use-and-function-calling

---

Topology Decision Tree

What is the task structure?
+- Linear pipeline (research -> draft -> review)?
|  -> Sequential / Pipeline
+- One coordinator delegates to specialists?
|  -> Supervisor / Worker (most common)
+- Peers swap control based on context?
|  -> Swarm / Handoff (OpenAI Swarm pattern)
+- Tree of sub-tasks?
|  -> Hierarchical (manager -> sub-managers -> workers)
+- Arbitrary directed graph with conditional edges?
|  -> Graph (LangGraph)
+- Independent agents across orgs?
   -> A2A protocol with shared task object

---

Topology Patterns

Supervisor / Worker (default)

A supervisor agent decomposes the task and routes each sub-task to a specialist worker. Worker results return to the supervisor, which decides the next step or finalizes.

• Pros: simple, auditable, easy to add workers
• Cons: supervisor is a bottleneck and a single point of prompt failure
• Frameworks: LangGraph supervisor, AutoGen GroupChat (with manager), CrewAI hierarchical

Swarm / Handoff

Each agent decides when to hand control to a peer by emitting a handoff(target_agent, context) tool call. No central supervisor.

• Pros: emergent routing, less prompt overhead per turn
• Cons: harder to debug, risk of ping-pong loops
• Frameworks: OpenAI Swarm / Agents SDK, Microsoft Agent Framework