Mental model
As agentic tasks grow in complexity, single-agent loops degrade due to prompt context clutter and tool distraction. The Multi-Agent Supervisor Pattern decomposes system capabilities into specialized, isolated worker subgraphs under a central supervisor router.
The supervisor acts as an executive manager:
- Receives complex user objectives.
- Evaluates current state and decides which specialist worker (e.g.,
ResearchAgent,CodeAgent,AuditorAgent) should run next. - Delegates sub-tasks and receives structured outputs through channel reducers.
- Determines when the overall objective is satisfied and emits the final answer.
Theory
Multi-agent coordination requires strict State Isolation:
- Parent Graph State: Holds global conversation context, overall progress, and supervisor routing decisions.
- Worker Subgraph State: Holds ephemeral scratchpad memory and specialized tool execution outputs.
- Message Handoffs: Workers communicate back to the supervisor via structured state updates, preventing internal worker scratchpads from inflating global prompt context windows.
Alternatives and trade-offs
Flat supervisor routing works best for 2-5 workers. Nested sub-supervisors (hierarchical teams) are required when domain sub-tasks exceed 10+ distinct tools. Peer-to-peer swarms pass control directly between agents without a central router, offering higher autonomy but lower auditability.
Failure modes and misconceptions
- Supervisor Ping-Pong: The supervisor repeatedly delegates between two worker agents without progressing toward completion. Mitigate with max iteration counts and explicit state flags.
- Leaky Subgraph Context: Passing entire conversation histories to specialized workers instead of minimal task summaries wastes tokens and introduces context noise.
Knowledge check
What is the primary architectural advantage of using a Multi-Agent Supervisor Pattern over a single prompt loop with dozens of tools?
Decision scenario
An enterprise customer support architecture routes incoming requests through a Supervisor LLM. The supervisor routes technical queries to a Diagnostics Subgraph and billing queries to a Payments Subgraph, maintaining separate scratchpads for each.
Learning outcomes
- Design supervisor routing topologies for complex multi-agent tasks.
- Isolate worker subgraph scratchpad state from global context windows.
- Evaluate trade-offs between supervisor routers and peer-to-peer agent swarms.
Trade-offs
Supervisor multi-agent topologies isolate prompt context and improve tool call accuracy, but increase token overhead for routing steps and require robust deadlock prevention.