All concepts
Multi-Agent Systems
Coordinate multiple specialized agents to solve a task collaboratively.
Agentic AI · Advanced · ~8 min
In plain English
Split the work between several specialists — a researcher, a writer, a reviewer — each with its own instructions and its own small set of tools.
Why it's worth your time
It's how you get past the point where one agent has too many tools and too much instruction to be reliable.
If you remember three things
- Specialization keeps each tool surface small
- Handoffs need a defined contract, not free-form chat
- Coordination cost is real and often exceeds the benefit
Overview
Decompose a complex task across several specialized agents — researcher, builder, reviewer — coordinated by a manager that assigns subtasks, tracks state, and resolves conflicts. Results are merged and reviewed before responding. Parallel specialization improves coverage at the cost of orchestration complexity.
How it works
- Start: Shared Task A complex task needs different skills or viewpoints.
- Shared Task -> Coordinator A manager assigns subtasks, tracks state, and resolves conflicts.
- Coordinator -> Research Agent One agent gathers evidence.
- Research Agent -> Builder Agent Another agent writes code, calls tools, or drafts output.
- Builder Agent -> Merge + Review Results are merged, deduplicated, and reviewed before response.
In an interview
Instead of one agent doing everything, you split work across role-specialized agents coordinated by an orchestrator that assigns subtasks and merges results. It helps when a task needs distinct skills or parallel exploration, but adds coordination overhead, token cost, and failure modes like agents duplicating work or disagreeing.
Production defaults
- Split when
- one agent exceeds ~8 tools or its system prompt covers unrelated jobs
- Topology
- one orchestrator plus specialists. Peer-to-peer chat between agents rarely converges
- Contract
- structured handoff objects, not conversational messages
- Budget
- a global step and token cap across ALL agents, not per agent
What breaks
- Agents talk to each other forever — No global budget. Cap total steps across the system, not per agent.
- Worse than the single agent it replaced — Coordination overhead exceeded specialization gains. Most tasks don't need multiple agents.