Multi-Agent Systems
Specialized agents that plan, code, and review in parallel.
One agent is a worker. Several agents with roles form a team — and often a better one.
Common patterns
Planner / executor
A planner breaks the goal into steps; an executor implements each step.
Planner: "1) add the field 2) migrate 3) update tests"
Executor: implements 1 → reports → implements 2 → …Coder / reviewer
A second agent reviews the first agent's diff, looking for bugs the writer is blind to. This catches a surprising number of real issues.
Specialist workers
Each agent gets a narrow job and a narrow toolset:
- Researcher — searches docs and summarizes.
- Implementer — writes code.
- Tester — writes/runs tests.
- Reviewer — critiques the diff.
Orchestrator
A supervisor routes work to specialists and merges results — the pattern behind frameworks like LangGraph, AutoGen, and CrewAI.
The cost
Every extra agent multiplies:
- Tokens (each agent re-reads context).
- Latency (they pass work back and forth).
- Failure surface (a bad handoff loses information).
When it's worth it
- Tasks with distinct phases (research → implement → verify).
- Large, parallelizable work.
- When an independent reviewer meaningfully improves quality.
When it's not
For a single, well-understood change, one good agent with a clean loop beats a committee. Start with one agent; add more only when a real bottleneck appears.
Multi-agent isn't "more is better." It's division of labor. Split the work at the points where one agent's context would overflow or its focus would drift.