BroadRouter
Get started

Multi-agent system

Updated

A multi-agent system is an LLM application in which several agents, each with its own prompt, tools and context, work on parts of one task. A common design is orchestrator-worker: a lead agent plans, starts worker agents in parallel and combines their results.

Also called: multi-agent, multi-agent orchestration, orchestrator-worker

One agent works through a task step by step. A multi-agent system splits the task and runs the pieces at once.

How it works

Anthropic describes its Research feature as an “orchestrator-worker pattern, where a lead agent coordinates the process while delegating to specialized subagents.” The lead reads the question, plans, starts several subagents that search in parallel, and writes the final answer from their summaries.

Other designs exist. The AutoGen paper (Wu et al., 2023) describes a framework in which “customizable, conversable” agents talk to each other, with humans and with tools, and developers program the conversation pattern in natural language or code.

A worked example of an orchestrator-worker run:

Lead:     plan -> start 3 workers
Worker 1: search docs for topic A  -> 1,500-token summary
Worker 2: search docs for topic B  -> 1,500-token summary
Worker 3: search docs for topic C  -> 1,500-token summary
Lead:     read 3 summaries -> final answer

Each worker spends its own tokens on searching and reading. The lead only sees the summaries.

Why it matters

  • Quality on broad tasks. Anthropic reports that a multi-agent setup with a Claude Opus 4 lead and Sonnet 4 workers “outperformed single-agent Claude Opus 4 by 90.2%” on its internal research eval.
  • Token cost. The same post says agents use about four times the tokens of chat, and multi-agent systems “about 15× more tokens than chats.” It also reports that token usage alone explained 80% of the variance in performance on BrowseComp. More agents usually means a bigger bill. The multi-agent cost post compares per-token and flat billing.
  • Fit. Anthropic notes coding is a weaker fit, since “most coding tasks involve fewer truly parallelizable tasks than research,” and tasks with heavy shared context or dependencies between agents suit it poorly.
  • Concurrency. Parallel workers are parallel requests, so your concurrency limit caps the speed-up.

On BroadNet Inference

BroadNet serves the models; the orchestration runs in your framework. Guides cover CrewAI and the OpenAI Agents SDK with a BroadNet base URL.

A pay-as-you-go key allows 30 requests per minute and 4 in flight. On the flat-rate plans for DeepSeek V4.1 Flash there is no token meter, which removes the token multiplier from the bill, but each session runs one request at a time. Buy one session for each model request that can be in flight at the same moment: three workers running at once need three sessions, and four if the lead keeps calling the model while they work. Running parallel coding agents on a flat-rate API walks through common patterns.

Questions

Is a multi-agent system always better than one agent?

No. Anthropic reports large gains on broad research tasks, but also that multi-agent systems use about 15 times the tokens of a chat, and that most coding tasks have fewer truly parallel parts than research.

What frameworks build multi-agent systems?

AutoGen, the OpenAI Agents SDK, CrewAI and coding tools such as Claude Code all support some form of it. AutoGen's paper describes agents that converse with each other, with humans and with tools.

How many parallel requests does a multi-agent system need?

One per agent that calls the model at the same moment. A lead plus four workers running together needs five requests in flight.

Read as MarkdownAsk ChatGPTAsk ClaudeAsk Perplexity