Multi-agents

A multi-agent system splits work across several LLM-driven agents instead of loading everything onto one. Each agent gets its own role, prompt, tools, and context window — a researcher that searches, a writer that drafts, a reviewer that critiques — and they coordinate by passing messages or handing off tasks. Common topologies include the supervisor (one orchestrator routes work to specialists), peer handoffs (agents transfer control directly), and parallel fan-out where subagents work independently and report back.

The honest reason to go multi-agent isn’t that it’s fancier — it’s context isolation and focus. One agent with twenty tools and a sprawling prompt gets confused; three agents with focused prompts and five tools each stay sharp. Subagents can burn thousands of tokens exploring a problem and return only a summary, keeping the main context clean. Specialization also means you can give each agent a different model — a cheap one for triage, a strong one for synthesis. The cost is real: more latency, more tokens, and coordination failure modes where agents duplicate work or lose shared state. Start with one agent; split when a single context demonstrably can’t hold the job.

In practice you’ll use a framework’s primitives — LangGraph supervisor graphs, CrewAI crews, OpenAI Agents SDK handoffs, or Claude’s subagents — or roll your own by exposing an agent as a tool: delegate_to_researcher(task) is just a function call that runs another loop and returns its final answer.