Scout7 logo

Scout7

Back

market_pov

Multi-Agent Systems: Stop Splitting The Work. Ask This One Question First.

September 11, 2026 · 9 min read · Scout7

Two famous posts came out one day apart and looked like opposite advice. They were not. Here is the one question to ask before you give one job to two AI agents.

Multi-Agent Systems: Stop Splitting The Work. Ask This One Question First.

Two famous posts came out one day apart in June 2025. They looked like opposite advice. They were not. Here is the one question to ask before you give one job to two agents.

Two posts, one day apart, and they were not really fighting

Should you give one job to two AI agents at once? Only when the second one does not need to see what the first one did.

People call two or more agents on one job a multi-agent system. That is all the words mean. This piece is about when multi-agent systems help you, and when they quietly waste your time.

The short version:

  • Ask one question before you split any job
  • The famous win came with a much bigger bill
  • Nothing looks broken when a split job goes wrong
  • Two agents help on separate work, not on shared decisions

On 12 June 2025, Walden Yan at Cognition published a post called "Don't Build Multi-Agents". His advice was simple. Do not hand one job to several agents at once.

About a day later, Anthropic published a post about doing exactly that. They said their setup beat a single agent by 90.2 percent on their own research tests.

One team said do not do it. The other team said it works. One day apart.

It looks like a fight. It is not.

Both posts land on the same rule. Before you give one job to two agents, ask one question. Does the second agent need to see what the first one did?

If yes, keep it in one thread. If no, split it.

That is the whole test. It takes five seconds.

The rest of this piece shows where that rule comes from, and what the famous 90.2 percent actually cost.

What the 90.2 percent really cost

What the 90.2 percent really cost

The Anthropic number is real. People quote it everywhere. Almost nobody reads the next few lines.

The June 2025 post from Anthropic engineering says a lead agent with worker agents beat a single agent by 90.2 percent on their own research tests.

In the same post they say two more things.

First, the split version used about 15 times more tokens than a normal chat. A token is a small piece of text. More tokens means the machine read and wrote a lot more.

Second, and this is the part that changes the story. They looked at what caused the gap. About 80 percent of it came down to one thing. How much each side read.

Read that again slowly. Most of the win came from reading more. Not from splitting the job.

There is a third thing sitting in the same post. The test was research work. Research is wide work. You can send ten questions out at once and nobody has to agree with anybody.

So the honest summary is this.

  • The win was real, and it was measured properly
  • The bill was large, about fifteen times a normal chat
  • Most of the gain came from the extra reading
  • The job was wide, which is the kind of job splitting suits

None of that says splitting is bad. It says the win was not free, and it was not magic. So the next question is why the same trick breaks on other work.

Why two agents break

Why two agents break

Here is the plain version. When two agents work on one job at the same time, neither can see what the other is doing.

Each one starts with an empty head.

Every guide to splitting work published this year tells you that is the good part. A fresh agent does not carry your long messy conversation with it. It just gets on with its own piece. That is true, and it is a real benefit.

It is also the whole problem.

Take one fact. It cannot see what the other one did.

That fact is a gift when the two parts have nothing to do with each other. That same fact is a disaster when the second part needs the first.

Same fact. Opposite result. Which one you get depends only on the job.

Walden Yan's post gives a real example. A team asked for a small phone game, a copy of Flappy Bird. The job was split in two. One worker came back with a background drawn in a Super Mario style. The other came back with a bird that did not match it.

Neither one failed. Both did good work. The two halves just did not go together.

Yan gives the reason. Every action an agent takes hides a decision inside it. Two agents make different decisions. You get a broken result.

And here is the part that catches people out. Nothing turns red. Both agents say done. You only find out at the end, when you try to put the pieces together.

Somebody counted what goes wrong in multi-agent systems

This is not just an argument between two companies. Someone went and counted.

A paper called "Why Do Multi-Agent LLM Systems Fail?" was presented at NeurIPS 2025, a research conference. The researchers read more than 1,600 recorded runs across 7 different tools, including AutoGen, ChatDev and CrewAI.

They sorted every failure into three groups.

  • The job was not described well enough: 41.77 percent
  • The agents did not agree with each other: 36.94 percent
  • Nobody checked the finished work: 21.30 percent

So more than a third of everything that went wrong was agents not agreeing with each other. That is not a small side problem. It is the second biggest cause on the list.

Across those 7 tools, the share of runs that failed ran from 41 percent to 86.7 percent.

That is not one broken tool. That is how these systems behave.

The work looks fine right up until the pieces have to fit together.

It also matches what the first two posts said. This is not about how big your company is. It is about the kind of job you handed over.

When splitting really is the right call

When splitting really is the right call

Splitting is not wrong. It is right for wide jobs, where nobody has to agree with anybody.

Here is what a wide job looks like.

  • Checking twelve websites at once for the same fact
  • Reading forty pages and writing one short note on each
  • Looking up ten prices, where no price depends on another price
  • Comparing five rivals, where each write up stands on its own

In all of those, no worker needs to see any other worker's answer. So the empty head costs you nothing at all. This is exactly the kind of work Anthropic was testing.

There is a second good reason to split. Sometimes the material simply will not fit in one go. Then splitting is not a preference. It is the only way.

That second reason is getting weaker every month. In September 2026 Anthropic shipped Fable 5.1. It holds one million tokens at a time. It also cut the price of re-reading stored text by 75 percent. More fits in one place now, and it costs less to keep it there.

And here is the part almost nobody quotes.

Anthropic named the limit themselves, in the same post as the 90.2 percent. Some work needs every agent to share the same background. Some work has one part leaning on another. They wrote that both kinds are not a good fit today. Then they added that most coding work is like that.

So the other side said the same thing, one day later.

Why this lands this month

Why this lands this month

Connecting agents to each other just got easy and official.

At Google Cloud Next 2026, the Agent2Agent protocol reached version 1.2. It is now run by the Linux Foundation, and Google says it is in real use at 150 organisations.

At the same time, the Salesforce 2026 Connectivity Benchmark asked 1,050 technology leaders across nine countries what they were running. The average company now has 12 AI agents. Half of them are not connected to anything.

So the tools to split work between agents are here. They are cheap, and they are easy to switch on.

The evidence about when splitting actually helps has not moved at all.

Easy does not mean right. It just means you can make the same mistake in more places now.

Frequently asked questions

When do multi-agent systems actually help?

They help most on wide jobs, where each part can run on its own. Checking many websites at once. Reading separate documents. Comparing rivals, where each write up stands alone.

Why do two agents fail on one job if both parts look fine?

Because the problem only shows up at the join. Each agent can do good work on its own. If they made different hidden decisions, the two parts will not fit together.

Did Anthropic prove that splitting work is better?

They showed a real gain on their own research test. But that setup used about 15 times more tokens than a normal chat. About 80 percent of the gain came from how much each side read, not from the splitting.

What is the one question to ask before splitting work?

Ask this. Does the second agent need to see what the first one did? If yes, keep it in one thread. If no, split it.

The only rule worth keeping

The only rule worth keeping

You do not need a diagram for this. You need one question, asked before you split.

Does the second agent need to see what the first one did?

If yes, keep it in one thread. If no, split it.

That question settles the fight we opened with. Cognition says do not split a job that hides decisions from itself. Anthropic shows a big win on wide research work. Then they say the leaning kind of work is not a good fit.

They agree. The bigger number just got shared more.

  • The job decides, not the size of your company
  • A fresh agent helps when the parts run on their own
  • A fresh agent hurts when part two leans on part one
  • Nothing turns red when it goes wrong, so you have to look

One more thing, and it is free. When you do split, check the join. Both halves can be good and the join can still be broken. The join is the only place the problem ever shows up.

Have you ever split one job between two agents? What did they disagree about?

References