The average score hid the break

Faster, and one thing quietly broke

Anthropic published an example from a sample customer-support agent. The team folded one specialist subagent into the main agent, and response time dropped. It also broke a process that had been passing. The main agent looked up the order and answered by itself, without asking the shipping specialist for the delivery date.

It broke in only one of two runs. Run it once, and you may see the good run and ship the change.

Why teams fold agents in the first place

The example comes from an Anthropic article published on October 22, 2026, about when to fold subagents into the main agent loop. Its claim: many enterprise agents work better with fewer subagents. One loop that loads skills and tools only when it needs them often gives better quality, lower cost and lower latency than a boss agent directing a team of specialists.

The reason teams build the other way is familiar. They draw the agent system to look like their org chart, which is Conway's law from 1968: systems copy the communication structure of the organization that builds them. Then handoffs drop details, and the boss agent collects rules that start to contradict each other. Anthropic reports one customer running about twice as fast on a single loop, and says tool choice tends to degrade once an agent has somewhere between 30 and 50 tool definitions.

This post is not about whether to fold. It is about how to find out.

It is not an agent problem

You can break something this way without writing any code. Edit the instructions of a Claude Project that answers HR questions, and the average answer improves while one policy question starts getting the wrong number of days. Swap a customer-service template for a shorter one, and most replies get warmer while one stops mentioning the refund window.

An average hides exactly this. The fix is boring and works.

A ten-question test sheet

Keep ten real questions that you already know the right answer to. Ten is enough to start. For an HR Project, take the ten questions employees actually ask, with the answer from the policy next to each. For an expense-classification prompt, take ten receipts whose category you are sure of.

Then follow five steps:

1. Run all ten before you change anything, and save the answers.
2. Change one thing.
3. Run all ten again, at least twice, because AI answers vary from run to run.
4. Compare question by question. Ask which tests flipped, not what the overall score is.
5. Keep the change only if nothing that mattered got worse.

Anthropic's rule for subagents is the same idea in one line: keep a subagent only if removing it scores worse. For a prompt, the version is to keep a paragraph of instructions only if cutting it makes some answer worse.

Subagents still earn their place in three cases: work that really runs in parallel, an independent check of another agent's output, and a single loop that degrades because there is too much in its context.

What the comparison looks like

You can ask Claude to do the tedious part. Paste the two sets of answers and ask it to line them up per question, say whether each answer changed, and judge whether the change is better, worse or just different. You read the flagged ones. Everything else you skip.

What to do this week

Pick the one Claude Project or prompt that people in your team rely on. Write down ten real questions with their right answers, run them twice, and keep the sheet. The next time someone says "I improved the instructions," you will have a way to find out.