multi-agent-patterns

Place a central supervisor agent in control that decomposes user objectives into subtasks and routes them to specialist workers. Implement a forward_message tool so sub-agents can pass responses directly to users, avoiding the telephone game problem where supervisor paraphrasing loses fidelity.

Supervisor patterns offer strict control and easier human oversight but risk bottlenecks and cascading failures. Swarm patterns allow direct agent-to-agent handoffs with no single point of failure, and benchmarks show swarms slightly outperform supervisors when sub-agents respond directly to users.

Use multi-agent designs when tasks decompose into parallel subtasks, require different tool sets or system prompts per subtask, or exceed single-context limits. Note that multi-agent systems consume roughly 15x baseline tokens, so upgrading the model may be a cheaper first step.

Use file system memory where agents read and write to persistent storage as the coordination mechanism. For simple subtasks, pass only instructions via function calls; reserve full context delegation for complex subtasks that genuinely need complete understanding.

Simple majority voting treats hallucinations as equal to sound reasoning, and agents exhibit sycophancy bias toward agreement. Use weighted voting by confidence or expertise, debate protocols with adversarial critique, and trigger-based intervention to detect stalls and mimicry.

LangGraph implements graph-based state machines with explicit nodes and edges, AutoGen uses conversational GroupChat patterns, and CrewAI provides role-based hierarchical crew processes. Each framework embodies different coordination philosophies suited to different workflow structures.