multi-agent-patterns

Create a central supervisor agent that decomposes tasks into subtasks, routes them to specialized workers, and aggregates results. Implement a forward_message tool so sub-agents can respond directly to users, avoiding the telephone game problem where supervisors lose fidelity when paraphrasing.

Supervisor patterns use a central agent for control and synthesis, while swarm patterns let agents hand off directly to each other via transfer functions. Swarms avoid single points of failure and can outperform supervisors when sub-agents respond directly to users.

Use multi-agent designs when single-agent context limits constrain task complexity, when tasks decompose into parallel subtasks, or when subtasks need different tools and system prompts. Note that multi-agent systems consume roughly 15 times more tokens than single-agent chat.

Yes, LangGraph implements supervisors as graph state machines with explicit nodes and edges. The references document shows a StateGraph implementation where a supervisor node routes to researcher and writer worker nodes based on task content.

Use weighted voting based on confidence and expertise rather than simple majority voting, or apply debate protocols where agents critique each other over multiple rounds. Monitor for sycophancy triggers where agents mimic each other without independent reasoning.

Common failures include supervisor bottlenecks from accumulated context, coordination overhead negating parallelization gains, agent divergence from objectives, and error propagation between agents. Mitigate with output schema constraints, checkpointing, convergence checks, and circuit breakers.