Glossary Entry

Multi-Agent System

An agent harness that runs several model loops, each with its own context window, on a shared task, such as an orchestrator that delegates to worker agents or peers that hand work to one another.

LLMs Architecture

Also called: multi-agent systems, multi-agent architecture, multi-agent architectures

Seed source: Anthropic

Splitting work across agents can add breadth, keep exploratory work out of the main agent’s context, and give a verifier a clean view of the result. Every boundary between agents is also a place where context gets summarized or lost, and agents writing to the same system can make conflicting decisions that none of them sees.

The evidence so far favours a narrow use: parallel agents for reading, searching, and reviewing, with writes kept in one place or in isolated workspaces behind a mechanical check. Multi-agent runs also typically cost several times the tokens of a single agent on the same task.