Multi-agent System

Like a team of specialists beats a generalist, specialized agents (one for research, one for code, one for analysis) produce better results than a single generalist agent.

Via protocols like MCP, structured messages, or shared memory spaces. Each agent has its role and tools.

A multi-agent system (MAS) is an architecture where several AI agents, each with its own role, tools, and memory, coordinate to solve a task that one agent would handle poorly. Instead of a single monolithic prompt, work is decomposed across specialized agents that exchange messages, delegate subtasks, and combine results. Frameworks like CrewAI, AutoGen, and LangGraph make this orchestration explicit and reproducible.

A common example is automated research: a planner agent breaks down a question, several worker agents search the web in parallel, and a writer agent synthesizes a cited report. Software pipelines work the same way, with a coder, a reviewer, and a tester agent. Tools like Claude Code, openclaw, and Relevance AI build real products on this orchestrator-worker pattern.

A single agent runs one loop, deciding and acting on its own with a fixed set of tools. A multi-agent system splits that work across several agents that specialize, run in parallel, and hand off tasks to each other. The trade-off: multi-agent setups scale to harder problems but add coordination overhead, latency, and more points of failure, so they pay off mainly on complex, decomposable tasks.

Coordination usually follows one of a few patterns. In orchestrator-worker, a lead agent plans and dispatches subtasks to workers, then merges their output. In agent handoff, control passes from one agent to the next as the task moves through stages. Some systems use hierarchical teams or a shared scratchpad. LangGraph models these as explicit graphs, while CrewAI frames them as roles in a crew.

No, not by default. ChatGPT is primarily a single conversational model, even when it calls tools or browses the web within one loop. It becomes part of a multi-agent system only when you orchestrate several model instances with distinct roles that delegate to and review each other. The LLM is the engine; the multi-agent system is the architecture wrapped around several such engines.

The most established options are CrewAI, which models agents as roles in a crew with tasks; AutoGen from Microsoft, which centers on conversational agents that message each other; and LangGraph, which represents agent workflows as explicit state graphs for fine control. OpenAI's Agents SDK and Relevance AI are also widely used. The right choice depends on how much you need explicit control versus quick orchestration.

Use one when a task is complex enough to split into distinct, parallelizable subtasks that benefit from specialized roles, such as research, coding plus review, or multi-step workflows spanning several tools. For simple, linear tasks, a single agent is cheaper, faster, and easier to debug. Reach for multiple agents only when coordination genuinely buys you better quality or throughput.