Glossary
multi-agent
Architectures where multiple LLM-driven agents collaborate or compete on a task, each with its own role, prompt, or specialization, coordinated by an orchestrator or message-passing protocol.
A system where two or more LLM agents communicate to solve a problem. Patterns vary: a planner agent splits work and delegates to specialist sub-agents; a debater pair argues toward a synthesis; a research lead spawns parallel investigators and aggregates their findings.
The empirical case for multi-agent is contested. Anthropic’s published research system uses one lead agent that spawns parallel sub-agents and reports gains over single-agent baselines on open-ended research tasks. Cognition’s June 2025 “Don’t Build Multi-Agents” essay argues the opposite: that fragmenting state across agents loses critical context and that single-agent designs are the safer default for software tasks.
Frameworks that target this shape: AutoGenagentsA Microsoft Research framework for multi-agent systems, with a conversation-pattern API for orchestrating multiple specialized agents to solve tasks collaboratively. Open full entry , CrewAI, LangGraph, AG2, MetaGPT. They differ on communication topology (group chat vs hub-spoke vs DAG) and on how shared state is represented. Whether multi-agent is genuinely useful or a wrapper around a more careful single-agent design is a live question through 2026.