Several model-driven agents working together, each with a defined role, handing work between them. One might plan, another write code, another review it. It suits problems that decompose cleanly and are large enough that a single context cannot hold everything. It also multiplies the number of ways things can go wrong.
Technical view
Common patterns include an orchestrator delegating to specialist workers, sequential pipelines with defined handoffs, and debate or critique arrangements where one agent reviews another’s output. The advantages are context isolation and parallelism; the costs are token consumption, latency and compounding error, since a mistake made early is inherited downstream. Reliability depends far more on clear interfaces and verification at each handoff than on the sophistication of any individual agent.
Related
Part of AI Terminology Index, section: Agents and Tools.