All writings
Essay · 4 min

One AI Agent Per Employee? Anthropic Says Think Twice

Anthropic's Nov 11, 2026 study found teams with one agent per person did worse than one shared agent. Here is what to do about it.

By Peter Walters

Anthropic published a study called "Worse Together" on November 11, 2026 on its Alignment Science blog. It found that when every person on a team has their own AI agent and all of those agents share one resource, such as a budget, a calendar or a group order, the group does worse than when one agent serves everyone.

What actually changed

Nothing shipped. This is research, and it is the first careful look I have seen at a setup many small teams are drifting into: each person with their own assistant, all touching the same stuff. The Anthropic post lists authors Sahan Paliskara, Nattaput Namchittai, Jacqueline Tay, Avery Griffin and Andrew Lampinen. The paper is on arXiv and the benchmark code, MAMUBench, is on GitHub.

They built four simulated environments: a shared API budget, a clinic calendar, a group food order, and a code merge queue with a release deadline. Then they compared one coordinating agent against one agent per user. Models tested included Claude Opus 5, Claude Sonnet 5, GPT-5.6-sol, GPT-5.6-terra and Qwen3.8-max.

The coordinator won in all four environments for both Claude models. On the shared API budget, Opus 5 scored 64 percent of the best possible outcome as a coordinator and 30 percent as a team of peers. On the group order, 88 percent versus 44 percent. Sonnet 5 showed the same pattern (75 versus 37 on the group order).

The failures are the interesting part. Agents stalled as teams grew. Some models cancelled other people's jobs without asking: 3.9 to 7.1 times per episode for the GPT and Qwen models, 0.59 for Sonnet 5, and zero for Opus 5. In the group order, a person's request, like a gluten-free requirement, often never made it into the context of the agent placing the final order. And in 58 percent of Opus 5 team runs and 53 percent of Sonnet 5 runs, agents made up claims about what other people wanted, versus 17 to 20 percent for a single agent.

Why it matters if you run a small team

If you have five staff and five separate AI assistants all booking, ordering or editing the same shared things, this is your future. It is happening now through tools that give every seat its own agent.

Here is my opinion. The study is simulated, with 25 scenarios per environment in the headline results, and some of the grading was done by other models. Treat the exact percentages as a signal, not a forecast for your studio. What I do trust is the shape: more agents with no referee means more stalling, more overwriting and more confident guessing.

What is overhyped is the idea that agents will "just coordinate with each other." Giving them a chat channel did not fix it.

What is underhyped is how cheap the fixes were. A designated team lead that ranked jobs closed the gap on the budget task. For the group order, a simple guard that blocked checkout until the final agent had read every pending message recovered 73 percent of failed Opus 5 runs and 62 percent of failed Sonnet 5 runs. That is a plumbing fix, not a smarter model.

What I would do with it this week

First, list the shared things in your business that more than one person (or agent) can change: the calendar, the ad budget, the website, the booking sheet, the price list. Those are your collision points.

Second, pick one owner for each. It can be a person or one agent. Everyone else asks the owner instead of editing directly.

Third, add a read-before-commit rule. Before anything changes a shared item, the agent has to read the latest notes and requests. If your tool cannot enforce that, put it in the instructions and spot check it.

Fourth, require a human yes for anything destructive. Cancelling, deleting and shrinking someone else's work should never be silent.

Fifth, tell agents to say "I don't know what Sam wants" instead of guessing. Fabrication was the failure that surprised me most.

When I built a daily brief agent for a real estate agent I work with, it served one person and read from one set of sources. I design the things I build for hotels and event spaces the same way: one agent holds the shared facts and rules, and everything else goes through it. I have not run a controlled test of this, so take it as how I approach the work, not as proof.

Sources

Want one practical AI move like this in your inbox each week? Join the letter, or see where to begin if you are just starting.

If this landed

I build websites, apps, and AI tools for people who can't stop thinking about a specific thing. If that's you, let's talk.

Book a conversation