Multi-Agent System

A multi-agent system decomposes a task across several LLM agents with distinct roles — a researcher, a coder, a reviewer, a planner — that communicate to produce a result no single prompt handles well. Common topologies include orchestrator–worker (a lead agent delegates subtasks), pipelines (each agent transforms the previous agent's output), and debate (agents critique each other before a verdict). The appeal is separation of concerns: each agent gets a focused system prompt, its own tools, and a smaller context, which usually beats one overloaded mega-prompt on complex work. The costs are real, though — token spend multiplies, latency stacks, and errors propagate between agents, so failures are harder to trace. Practical guidance for SaaS teams: start with one agent, split only when a role demonstrably degrades quality, and add tracing plus per-agent evals so you can see which link in the chain broke.