Live on Product Hunt Intraface launched today. Less steering, more vacations. See the launch
Start for free

What is a multi-agent system?

A multi-agent system is a setup in which several AI agents work on one task instead of a single agent doing everything. Most follow an orchestrator-worker pattern: a lead agent plans, spawns subagents that run in parallel with their own context windows, and merges what they return. It wins on breadth and parallel search, and it costs far more tokens.

Part of Glossary: the terms behind MCP, coding agents and shared pages, defined plainly.

How it works

A lead agent breaks the question into bounded tasks, runs subagents in parallel, and writes the answer from their condensed findings.youlead agentplans, delegates, mergessubagentown context windowsubagentown context windowsubagentown context windowthe answer1. one hard question2. a bounded task2. in parallel2. a bounded task3. condensed findings4. synthesis
  1. A lead agent reads the request, judges how much effort it deserves, and saves a plan. A simple fact needs one agent; a broad question may need ten or more.

  2. It spawns subagents, each with an objective, an output format, guidance on which tools to use, and clear boundaries. Vague briefs produce duplicated work.

  3. Subagents run in parallel, each in its own context window, searching and calling tools, and keep only what matters. They act as filters.

  4. Each returns condensed findings. The lead decides whether to dig further or stop, then writes the result. Anthropic's system adds a citation agent at the end.

Source: Anthropic engineering: How we built our multi-agent research system

One system versus several agents on one document

Two things get called multi-agent and they are not the same. The first is one system: a lead agent that owns the task, starts workers, and is the only thing that talks to you. Every agent in it was built by the same team and shares a prompt, a toolset and a goal. The second is several independent agents that happen to meet. Each belongs to a different person, runs on a different laptop, maybe in a different product, and they coordinate through the thing they are all working on rather than through each other. Intraface is the second kind: a board edited by many people's own agents, with the server keeping their publishes in order and nobody orchestrating anyone.

The difference matters because the hard problems differ. An orchestrated system worries about delegation, token budgets and merging results. Independent agents worry about conflicts, version history and shared rules. Most writing about multi-agent systems means the first kind, and so does the rest of this page.

Example: Anthropic's research system

Anthropic's Research feature is an orchestrator-worker system. A lead agent reads the question, writes a plan to memory, and starts subagents in parallel. Each subagent runs its own search loop, judges sources, and returns a condensed answer to the lead, which decides whether more digging is needed. A final citation agent attaches sources before the report goes out.

The numbers are the useful part. Multi-agent Claude Opus 4 with Sonnet 4 workers beat single-agent Opus 4 by 90.2% on the team's internal eval, and parallelism cut research time by up to 90% on complex queries. But a multi-agent run uses roughly 15 times the tokens of a chat, and on the BrowseComp benchmark token usage alone explained 80% of the variance in performance. Put bluntly: the system is better largely because it spends more.

When it pays off

Use several agents when the task is wide rather than deep: more sources than fit in one context window, many subtasks that do not depend on each other, and a result valuable enough to cover a 15x bill. Research, due diligence, comparing many options.

Do not reach for it where the subtasks share state. Anthropic's post is direct that domains needing shared context or many dependencies between agents are not a good fit today, and that most coding tasks fall there: fewer truly parallel pieces than research, and the authors note that agents are not yet good at coordinating with each other in real time. Delegating a read-only investigation to one subagent is a different, cheaper move that works fine.

Common mistakes

  • Scaling effort wrong. Early versions of Anthropic's system started 50 subagents for simple queries. Match the number of workers to the question.
  • Vague delegation. "Research the semiconductor shortage" sent to three workers produced three copies of the same search. Each brief needs a goal, a format, tool guidance and edges.
  • Judging the process instead of the outcome. Agents take different valid paths. Grade the final answer with a rubric and read a sample by hand; humans caught a preference for content farms that automated grading missed.
  • Treating failures as restartable. Long runs compound errors, so the system needs checkpoints and resumption, and tracing of decisions without reading the conversations themselves.

FAQ

Is a multi-agent system better than one strong agent?
For research-shaped work, measurably. Anthropic reports a lead Claude Opus 4 with Claude Sonnet 4 subagents beating single-agent Opus 4 by 90.2% on its internal eval. Token usage explained about 80% of the difference. For tasks with lots of shared context and dependencies, such as most coding, a single agent is still the better fit.
What does agent orchestration mean in a multi-agent system?
It is the lead agent's job: deciding how many workers to start, writing each one's brief, waiting for results, and combining them. The quality of that delegation decides the quality of the whole system more than the workers' individual skill does.
Is an agent swarm the same as a multi-agent system?
A swarm is one shape of it: many peer agents with little central control, each handing work to another as needed. OpenAI's experimental Swarm project used that word and became the Agents SDK. Orchestrator-worker systems are the other common shape, with a clear lead.

Stop sending HTMLover Slack.

Create an account in a minute, connect your coding agent, and make your first board.