Claude Agent Teams vs One Agent: Why Anthropic Says Fold Your Subagents (We Tested It)

Many AI agent teams feel slower than one good chat. A planner, a writer, a checker and a boss on top, all passing notes. On October 22, 2026, Anthropic told builders to fold most of those agents back into one, and we tested it: a 6-agent team and one Claude Code agent built the same app with the same score, and the single agent finished about 6 times faster.

Anthropic's own test kit backs this up. One fold made an agent 29% faster with 24% fewer model calls, and the score went up. But it also broke the simplest question a customer asked. We'll get to that.

100 agents do the work of 9

A hundred AI agents working together get about as much done as nine. Not a hundred. Nine. That number comes from the Oxford philosopher Toby Ord, who borrowed a rule from economists. When N workers share one task, they get about as much done as one worker going N to the power of lambda times as long.

Lambda is the stepping-on-toes number. If lambda is 1, a hundred workers do a hundred times the work. On Terminal-Bench, a coding benchmark, his estimate for AI agents is 0.48. So a hundred agents get roughly nine agents' worth done. That's about the same drop-off economists see in human teams. You wouldn't put twenty people on one small feature. Don't do it with agents either.

What a subagent really is

A subagent is a helper in its own room. It has its own notebook, its own tools and its own instructions. When it's done, it can't hand you the notebook. It slides a short note under the door. That note is the whole problem. And sometimes, it's the whole point.

Ana's order: the game of telephone

Meet Ana. She types one question into a store's chat: "Where's my order 5002?" Behind that chat box there's no single assistant. There's a router. It sends Ana's question to the orders agent, which looks it up and writes a note: "shipped." The router reads the note, then asks the shipping agent for a date. Shipping writes its own note.

Every handoff is a game of telephone. Every note is a little shorter than what the last agent actually knew. And every hop is another model call Ana sits and waits through.

Why your agents copy your org chart

Companies build it this way because of an idea from the 1960s called Conway's law: your systems copy the shape of your org chart. A shipping team builds a shipping agent. An orders team builds an orders agent. Six teams, six agents, whether the problem splits six ways or not. Anthropic's Applied AI team says they keep seeing exactly this.

5 ways AI agent teams fail

1. Lost context. Details get dropped between agents.
2. Sprawl. A refunds agent, a returns agent, then a refunds-and-returns agent, because nobody checked. Some companies end up with hundreds that overlap.
3. The Frankenprompt. Every team adds one sentence to the system prompt until the rules fight each other and the agent has no single voice.
4. Nobody tests the whole thing. Each team grades only its own piece.
5. Speed. One customer compared their multi-agent setup with a single loop. The single loop ran about twice as fast.

The fix: one agent, skills and tool search

The alternative is one agent, one context window (its working memory), running in a loop. Picture one chef instead of six cooks passing plates.

Above the chef is a shelf of recipe books, and the chef only sees the titles on the spines. When an order needs the returns recipe, the chef pulls that one book down. That's a skill. Under the counter is a drawer of labeled tools. The chef doesn't dump the drawer out; they search for the one they need. That's tool search.

Anthropic calls this progressive disclosure: short descriptions up front, full detail only when it's needed. It's how they cut Claude Code's own system prompt by more than 80% for the newer models.

Why it works now

• Shared context. A chef who saw every step can connect them. A pile of notes can't.
• No bloat. The books stay closed until they're needed.
• Better models. Since Opus 4.5, models handle far more in one loop than most teams expected. Give a newer model a tool that checks refund status, and it checks the refund went through without you writing a rule telling it to.

Anthropic says rigid multi-agent setups were often the very thing stopping a new model from doing more of the work.

When to keep a subagent

This is not "delete all your subagents." Anthropic still uses them, including in its own research system. Keep one when it passes any of three tests:

1. Your single loop shows a measurable drop from too much context. Measurable means a test score went down, not a feeling.
2. You need several jobs done in parallel, each in a fresh window you throw away after.
3. One agent keeps grabbing the wrong tool, or your rules really do conflict.

When you do split, split by what each helper needs to know, not by job title. A reviewer should see only the plan and the finished work, so it can't inherit the builder's blind spots. Then ask one question: would what the main agent already knows help this task, hurt it, or not matter? If it helps, keep the task in the loop. If it hurts or doesn't matter, send it to a fresh window.

Our test: 6-agent team vs 1 agent

We didn't want to just take their word for it. Same job, two ways: a booking page for a made-up hair salon, built from the same one-page spec.

• Lane A: an AI team of six subagents (planner, backend builder, frontend builder, security checker, tester, reviewer) plus a coordinator told not to write code.
• Lane B: one Claude Code session builds the whole thing, then one fresh reviewer sees only the spec and the code.

A hidden test neither lane saw tried 10 things a real customer would do: book, the 5 PM cutoff, closed Sundays, double booking, a beard trim only one stylist offers, phone formats, cancel frees the slot, no cancelling within 24 hours, reschedule, and booking from two other time zones. Each lane ran twice.

Same quality. But the team took about 6 times longer and burned about 9 times the tokens. Nobody lost a detail this time. They just spent the time passing notes. The cost column is Claude Code's reported API-equivalent figure; on a subscription plan it comes out of your usage limits, not a bill.

The catch: what broke

Back to Ana. In Anthropic's own recorded run, they folded the returns agent into the main loop. It got faster and cheaper, the score went from 92% to 96%, and one task that used to fail started passing.

But Ana's question went from passing two runs out of two to passing just one. The main agent was now holding the order-lookup tool, so it answered on its own, said "shipped," and never asked for the delivery date. That break showed up in only one run out of two, which is why their rule is: run every task more than once.

In a second run they kept a different agent but called it as a tool. It was 8% faster, and the score slipped from 94% to 91%. Folding isn't magic. It's a test. You keep whatever scores better.

The 4-step fold playbook

1. Write ten golden test tasks. Real questions, with what "done" means for each. Ten is enough to start.
2. Merge agents that always run together, like orders and shipping.
3. Call the agents you keep as tools, so your main agent keeps talking to the customer. That's your baseline.
4. Fold one agent at a time. Its prompt becomes a skill and its API calls become tools. Don't just look at the final score; look at which tasks flipped.

Group your tools by what the user wants, not one tool per API, because past about 30 to 50 tools the agent starts picking the wrong one. Anthropic ships a free kit that walks Claude Code through this audit, with a no-cost dry run you can try in ten minutes. The full write-up is in Anthropic's guide, When to fold subagents into your main agent loop.

Not a coder? Same rule

If you copy-paste between five custom chatbots, one to plan, one to write, one to edit, try one chat with a saved skill that holds your steps. Then compare the two on ten real tasks.

Get the Fold Checklist

With one agent, Ana gets her answer in one go: "It's on its way, arriving October third." No telephone. We put the whole method on one page: the three keep-tests, the help, hurt or doesn't-matter question, the four-step fold and a ten-task template. Comment FOLD on the video, or grab it below.

How we tested: Claude Code 2.1.310, headless, model Opus 5.5, October 26, 2026. Same one-page spec for both lanes, a hidden 10-flow browser test, two runs per lane, wall-clock minutes. One small test, not a study. Independent, not sponsored by or affiliated with Anthropic.