One customer measured a single agent loop against its multi-agent harness. Roughly 2x faster, Anthropic's Applied AI team reports. I run helper agents every day. This one is aimed at me.

I read it looking for the line that says I am wrong.

The headline will be "multi-agent is over."

The real finding is who designs all those agents, and why.

What Anthropic wrote (Diane Sarkis, Applied AI, Oct 22):

→ One company found "the dozen of specialist subagents within its customer-facing agent" made on-brand consistency hard and added latency.

→ Anthropic's usual feedback on designs like that: "one clear system prompt, 4-5 thin data tools, and firm output guardrails would cover most of the use case."

→ A different customer tested its multi-agent harness against a single agent loop and "saw a roughly 2x faster runtime with the latter".

→ Anthropic "reduced Claude Code's system prompt for more advanced models by over 80%".

→ The ceiling is tools, not agents: "once you exceed 30-50 tool definitions in the system prompt, tool choice tends to degrade."

What the headline hides:

The piece names where a dozen subagents come from: "the agent architecture becoming a copy of their org chart."

Then the prompt rots the same way: "Every team adds a sentence to the system prompt, resulting in conflicting rules and no consistent voice."

Twelve subagents is what happens when twelve teams each want one.

I run a one-person company on AI agents, so I have no org chart to copy. Still, my daily runs use one main agent for the judgment and the clicking, and hand the bulk reading to short-lived helpers with a fresh context window. They return a conclusion under 40 lines, never the page. One browser helper at a time.

That is one of Anthropic's three cases that still earn a subagent. I have not measured the speed. I keep them because the main agent's context stays clean, and that I can see.

If you run agents, this is the test. Anthropic keeps subagents only when:

Rule 1: "A single agent loop hits measurable performance degradation from context bloat". Measurable. If you have not measured it, you do not have it.

Rule 2: "Multiple tasks need to be tackled in parallel with fresh context windows". Reading is parallel. Deciding is not.

Rule 3: "A single agent struggles to select appropriate tools or there are conflicting instructions in the system prompt". First fix: delete tools, not add agents.

Fewer agents, thinner tools and a stronger model is good engineering news. I mean that.

You did not build a multi-agent system. You gave your org chart an API key.

If you run agents: how many subagents are in your loop, and which of the three conditions earns each one? One line is enough. 👇