When Do Several AI Agents Work Better Than One?

Imagine the Lagos fabric shop from the first post wants a price report every Monday morning: what ten rival online sellers charge for the same ankara print. The shop’s developer could build one agent that visits the sellers one after another. Or they could build a lead agent that gives each seller to its own helper and combines what the helpers find. Which is better? The honest answer starts with another question: does this job need a team at all?

(This is an illustrative scene, not a real business.)

What is a multi-agent system?

A multi-agent system is a set of AI agents that work together on one job. Anthropic’s engineering team, in a post of 13 June 2025 about its Research feature, wrote that “A multi-agent system consists of multiple agents (LLMs autonomously using tools in a loop) working together.” We saw in the second post that one agent is a model running in a loop with tools. A multi-agent system runs several of these loops. Each has its own instructions, its own tools and its own context window, which is everything the model can look at when it writes one response.

Should you start with several agents?

No. OpenAI’s guide, as of October 2026, says: “Our general recommendation is to maximize a single agent’s capabilities first.” It adds that more agents “can introduce additional complexity and overhead”. Anthropic’s December 2024 post “Building effective agents” gives related advice in general terms: “finding the simplest solution possible, and only increasing complexity when needed”. Anthropic has since noted that much of the tooling has changed, so we use only its general guidance. The fourth post in this series covers that choice.

OpenAI’s guide, as of October 2026, does name signs that it may be time to split. One is agents that “fail to follow complicated instructions or consistently select incorrect tools”. Another is prompts full of if-then-else branches that are hard to scale.

What are the common ways to arrange several agents?

OpenAI’s guide, as of October 2026, describes two broad patterns.

• Manager pattern. “A central ‘manager’ agent coordinates multiple specialized agents via tool calls.” OpenAI says it suits workflows where “you only want one agent to control workflow execution and have access to the user”.
• Decentralized pattern. “Multiple agents operate as peers, handing off tasks to one another based on their specializations.” A handoff, in OpenAI’s words, is “a one way transfer”: the next agent takes over the conversation.

Anthropic’s December 2024 post describes a similar design that it calls orchestrator-workers, in which “a central LLM dynamically breaks down tasks, delegates them to worker LLMs, and synthesizes their results.” In our reading, this is close to OpenAI’s manager pattern. An orchestrator is the agent in charge: it splits the work, hands out the pieces and combines the results. That is our plain-words definition, not a quotation.

Illustrative example. Imagine the shop’s customer assistant. In a manager design, it is the only agent that talks to the customer. When asked about a return, it calls a separate returns agent like a tool, reads the answer and replies itself. In a decentralized design, a front-desk agent decides the question is about returns and hands the whole conversation to the returns agent, which then talks to the customer directly.

When did several agents help in a real system?

Anthropic’s Research feature is a well-documented case. In its June 2025 post, a lead agent plans the work, then creates helper agents, which Anthropic calls subagents, that search different parts of a question in parallel. Each subagent has its own context window and condenses what it finds down to the most important parts for the lead agent, which combines them. Anthropic says its internal tests show that such systems “excel especially for breadth-first queries that involve pursuing multiple independent directions simultaneously”.

Anthropic reported that a system with Claude Opus 4 as the lead and Claude Sonnet 4 as subagents “outperformed single-agent Claude Opus 4 by 90.2%” on its internal research evaluation. That is Anthropic’s own test, on its own task, with 2025 models. We have not seen or repeated it, so read it as one company’s result, not a promise for your job. Anthropic also said that running helpers, and the tools they use, in parallel cut research time by up to 90 percent for complex queries.

Why does it help? Anthropic’s answer is plain: “Multi-agent systems work mainly because they help spend enough tokens to solve the problem.” A token is a small piece of text, often a word or part of a word. In one analysis Anthropic reported that token usage by itself explained 80 percent of the variance in the BrowseComp test, which measures how well a browsing agent finds hard-to-find information. In our reading, part of the gain from a team is simply more work done in total.

What does a team of agents cost?

A lot more. Anthropic wrote in June 2025: “In our data, agents typically use about 4× more tokens than chat interactions, and multi-agent systems use about 15× more tokens than chats.” It concluded that “For economic viability, multi-agent systems require tasks where the value of the task is high enough to pay for the increased performance.” Those are Anthropic’s figures for its own systems at the time. Yours will differ, so measure.

What can go wrong with a team of agents?

Anthropic listed early failures of its own system: “spawning 50 subagents for simple queries, scouring the web endlessly for nonexistent sources, and distracting each other with excessive updates”. Vague delegation was another problem. A lead agent told a helper to “research the semiconductor shortage”, and Anthropic reported that one subagent looked at the 2021 automotive chip crisis while two others duplicated each other’s work on 2025 supply chains. Its fix was to give each subagent “an objective, an output format, guidance on the tools and sources to use, and clear task boundaries”.

Some jobs are a poor fit. As of June 2025, Anthropic said that domains in which all agents must share the same context, or have many dependencies between them, “are not a good fit for multi-agent systems today”. It gave most coding tasks as an example, because they “involve fewer truly parallelizable tasks than research”, and said agents “are not yet great at coordinating and delegating to other agents in real time”.

Debugging is harder too. Anthropic wrote that agents “are non-deterministic between runs, even with identical prompts”. In our reading, every extra agent is another place where an instruction can be misread or a result lost, so a team needs more checking, not less. The next post covers how to test an agent.

How do agents from different companies talk to each other?

Everything above is about agents built by one team inside one system. A different problem is an agent from one organization working with an agent built by someone else. The Agent2Agent (A2A) protocol is an open standard for this. A protocol is an agreed set of rules for how two programs talk. The A2A project’s own pages, as of October 2026, describe it as “an open standard for communication between AI agents” that gives agents “a common language, whatever framework or vendor built them”.

In A2A, each agent publishes an Agent Card, which the project describes as a JSON document that acts as a “digital business card”. It describes the agent’s identity, capabilities, service address, skills and how to authenticate. Another agent reads the card, then sends it work. Longer jobs become tasks, which A2A defines as “a stateful unit of work” with a defined lifecycle. The agents stay opaque to each other: the project says they collaborate “without exposing their internal logic, memory, or proprietary tools”.

How does this relate to the Model Context Protocol from the sixth post? The A2A project’s documentation, as of October 2026, says MCP is “for agent-to-tool communication” and A2A is “for agent-to-agent communication”, and that the two “are highly complementary”. It also says A2A is “Not a sub-agent or tool-call protocol”. In our reading, the lead-and-helpers design from the Research example does not need A2A. A2A is for agents that live in separate systems.

As of October 2026, the project’s home page says A2A was originally developed by Google and donated to the Linux Foundation. A project post of 27 August 2026 says it has been accepted at the Agentic AI Foundation, which also hosts MCP. The same post says over 150 organizations back it. That figure is the project’s own, and we have not checked it.

Illustrative example. Imagine the courier also runs an agent. The shop’s agent could ask it for a delivery update through A2A, without seeing how the courier’s agent works inside. Neither agent is real. In our reading, most small businesses will not need this soon.

So how do you decide?

In our reading, ask four questions in order.

1. Does one agent with clear, well-described tools already do the job? If yes, stop. OpenAI says to start there, and Anthropic says to find the simplest solution that works.
2. Does the work split into parts that can run independently, at the same time? Breadth, such as ten sellers or ten documents, fits. A chain of steps that each depend on the last does not.
3. Is the result worth several times the cost? Anthropic’s multiples suggest planning for that.
4. Can you check the output? A team that cannot be tested is a bigger risk than a single agent that cannot be tested.

For the shop’s Monday report, our reading is to begin with one agent and a step limit, as in the eighth post. The developer should move to helpers only if one agent proves too slow or loses track of the sellers.

What to remember

• Start with one agent. Add more only when a single agent, with clear tools, fails at a job that splits naturally into parts.
• Teams of agents can do more on breadth-first jobs, but they use many more tokens and fail in ways that are harder to see. Anthropic’s figures are from June 2025 and are its own.
• MCP connects an agent to tools. A2A connects separate agents to each other. Helpers inside one system do not need A2A.

If you are weighing automation against agents for your own business, our guide to workflow automation with AI shows simpler options. To revisit earlier posts, see Do You Need an AI Agent, or Is a Simple Automation Enough? and How Do AI Agents Remember Things? Context Windows, Memory and RAG.

Next in the series: how to test whether an AI agent is any good.