Multi-Agent Systems: Getting Several AIs to Work as a Team

Four orchestration patterns, Anthropic's Research system as a case study, how to brief sub-agents, and when a single agent is the better choice.

The marketing team at a skincare brand wants a report comparing 5 competitors: prices, promotions, target customers and what people say on social media. Give the whole job to a single AI Agent and it has to research one brand at a time, keep everything in one context, and that context keeps growing until it starts to forget what it found at the beginning. This is the point where many teams start looking at Multi-Agent systems.

What is a Multi-Agent system

A Multi-Agent system has several AI Agents, each with its own role, tools and context, working together under a designed coordination pattern called orchestration. If you are not yet sure how a single agent works, read The Agent Loop first and this article will be easier to follow.

Four common coordination patterns

Anthropic's Building effective agents summarises several patterns used in practice. The first three below come from that article; the fourth is a hand-off pattern that is common in real deployments.

1. Orchestrator-Workers

One lead agent breaks the task down, delegates pieces to worker agents, then combines the results. It suits tasks where you cannot predict in advance how many parts the work will need.

2. Parallelization

Independent subtasks run at the same time (sectioning), or several agents do the same task and vote (voting) to increase confidence.

3. Evaluator-Optimizer

One agent produces work, another critiques it and sends it back for revision, looping until it meets the bar. It suits writing or translation with clear quality criteria.

4. Handoff

Work passes between agents with different specialities. For example, a customer intake agent hands over to a claims agent once it realises the case is about a damaged product.

Case study: Anthropic's Research system

In an engineering post from June 2025, Anthropic described how Claude's Research feature uses the orchestrator-workers pattern: a lead agent plans and spawns several subagents that search in parallel. Several reported results stand out:

• The multi-agent system (Claude Opus 4 as lead, Claude Sonnet 4 as subagents) outperformed a single Claude Opus 4 by 90.2% on the company's internal research eval
• But it used about 15 times the tokens of a normal chat
• Running subagents and tools in parallel cut research time for complex queries by up to about 90%
• Early versions spawned as many as 50 subagents for simple queries, so the team had to write rules that scale the number of agents to the difficulty of the task

The key takeaway: multi-agent helps most when the work is "broad" and can be split in parallel, but it comes at a much higher cost.

The heart of delegation: a good brief

Anthropic found that vague instructions led subagents to duplicate work or leave gaps. A brief for a worker agent should include four things:

1. A clear objective for that subtask
2. The output format to return, such as a 5-column price table
3. Guidance on tools and sources to use
4. Task boundaries: what not to do

Back to the skincare report. The lead might assign one subagent per brand, each with a brief like "Collect prices for the 3 best-selling products, this month's promotions and the issues customers mention most. Return a table. Do not write an overall analysis." Then another agent compiles the report, and one more checks that every number has a source.

When a single agent is better

Anthropic notes that multi-agent designs fit poorly when every agent needs the same context or when steps depend heavily on each other, and that most coding tasks have fewer parallelisable parts than research. The paper Why Do Multi-Agent LLM Systems Fail? (Cemri et al., 2025) also observes that performance gains on popular benchmarks are often minimal. So start with one agent and split only when you hit a real bottleneck.

Watch out for

• Cost: multiplying tokens has to be worth the value of the task. See how to keep costs down in Cutting AI Agent Cost and Latency
• Information dropped along the way: one agent may not pass key information to another, or may ignore a teammate's output
• Harder debugging: you need logs for every agent, or you will not know where it broke. Read more in Why Agents Fail

Adding agents does not add intelligence. It adds team management.

What to do next

1. Get a single agent working well first, and note where it gets stuck
2. If it struggles because the context is too long or the task too broad, split out subtasks that can run in parallel
3. Write each agent's brief with all four elements
4. Add a final checking agent before results reach people
5. Measure against the single-agent version on both quality and cost

Want to wire up a multi-agent workflow yourself? Try the course Build an AI Automation Agent with Make and n8n, and if you are interested in roles that use this kind of AI system design, browse open positions on Plants.work Jobs.