Multi-Agent System

A multi-agent system is an AI architecture where two or more agents work together to accomplish a goal that would be difficult or impossible for a single agent. Each agent has a specific role, and they coordinate through defined handoff protocols or an orchestrator agent.

Why multi-agent systems exist

Single agents have limits: - Context windows fill up on complex tasks, causing the agent to forget earlier instructions - Quality drops when one agent tries to do too many different things - Speed is constrained because a single agent processes tasks sequentially - Tool access varies — different tasks may need different tools or permissions

Multi-agent systems solve these by distributing work across specialists, each operating within their own context and area of expertise.

Coordination strategies

There are three primary ways to coordinate agents:

Sequential (pipeline): Agent A finishes, passes output to Agent B, who passes to Agent C. Best for workflows where each step depends on the previous one. Simple to build, easy to debug, but slow.

Parallel (fan-out/fan-in): Multiple agents work simultaneously on independent tasks. A coordinator collects all results when they finish. Best for research, data processing, and any task where sub-parts are independent. Fast, but requires merging results.

Hierarchical (orchestrated): A manager agent delegates to workers, who may delegate to their own sub-agents. Best for complex projects with mixed task types. Flexible, but adds coordination overhead.

Cost reality

Multi-agent systems multiply token usage. A system with 5 agents typically uses 3-5x the tokens of a single agent. Smart cost management includes: - Using cheaper models (e.g., Haiku) for simple sub-tasks - Reserving expensive models (e.g., Opus) for the orchestrator and quality-critical steps - Setting token budgets per agent to prevent runaway costs

When NOT to use multi-agent

Most tasks do not need multiple agents. Use multi-agent systems only when you hit specific limits of single agents. If the task fits in one context window, one agent does it well, and speed is not critical — a single well-prompted agent is cheaper, faster, and easier to debug.

The right number of agents is the minimum needed to do the job well.