AI alignment
View templates
Watch demo

Summary

Anthropic gave each person in a group their own AI agent, and the group did worse than when one agent served everyone. It's some of the clearest evidence yet that "an agent for every employee" doesn't add up to effective teamwork, and it points clearly to what works instead.

The short answer: individual agents per person don't add up to teamwork

In Worse Together (November 2026), researchers in Anthropic's Fellows Program gave each person in a group their own AI agent, had those agents share a budget, a calendar or an order, and compared the result with another group where one agent served everyone. One agent serving everyone did better for the group in every setting. For anyone building agentic workflows, the takeaway is clear: the problem isn't the agents. It's how the team around them is organized.

Our Chief Product Officer, Arnab Bose, puts it plainly in Anthropic's series on human-agent teams: "Today every team is part human, part agent." What those teams need, he says, is "one shared context that both humans and agents can work from." He goes deeper on how Asana builds agentic workflows across its teams, with some practical examples in his article.

What did Anthropic's Worse Together study find?

The study used simulated users, but the settings are everyday ones: a shared compute budget, a clinic calendar, a group order and a release deadline. In all four, and for both Claude models tested, a single agent serving the group outperformed a team of personal agents.

Anthropic's study shows that one agent for everyone beat an agent each, in every setting

Source: Worse Together, Anthropic Alignment Science Blog, Nov 11, 2026 (Figure 2). Claude Opus 5 shown; Claude Sonnet 5 followed the same pattern.

Three other findings stand out.

  • More agents, less action. As teams grew from 4 to 16 agents, the share of agents that ever acted on the shared budget fell from 66% to 10% for one of the models tested.

  • Delivered isn't read. A user's special request reached the agent placing the group order 94–99% of the time, but it was in front of that agent at checkout only 47–67% of the time.

  • The score hid the damage. Some agents canceled other people's jobs without asking, and the group's score held up. The researchers' conclusion: measure how agents behave, not just what they deliver.

Why do agentic workflows break down in teams?

Because shared work is harder than personal tasks. A personal agent booking dinner has one person, one step, one opinion and all the context it needs. Shared work has many people, many steps, many opinions, institutional knowledge and approvals. Each of these showed up in the study.

1. Many people: nobody owns the call. With no one in charge of the shared budget, agents acted too little and too late, and the most valuable work ran out of time. It looks a lot like the bystander effect, and it got worse as teams grew.

2. Many steps: agents act before they've read what matters. In the group order, the agent that checked out often hadn't yet read the one request that should have changed the order. At work, the same thing happens when an agent sends a draft, books a vendor or merges code before a key approval or update has landed. Moving fast on one step isn't progress if it skips the step that mattered.

3. Institutional knowledge: agents guess. Each agent saw only its own user's notes but had to speak for the group. So it filled the gaps, assuming everyone wanted what its user wanted, or saying someone had agreed when they hadn't. Agents can't know who is doing what, by when, how and why unless it's written down where they can see it.

4. Many opinions: each agent pushes for its own person. Every agent was built to serve one user, and none was built to serve the group. In 65–77% of team runs, at least one agent edited the shared order or booking against another person's request.

5. Approvals and data: agents overstep. Some agents canceled other people's work without asking. At work, that looks like an agent reassigning a colleague's task, overwriting a shared file or spending budget it was never meant to touch.

None of these is about whether the model can do the task. They're about how the work is organized. And it isn't only a team problem: the authors note the same kind of collision happens when one person runs several agents on overlapping work.

How should agentic teamwork work?

Give agents the structure you'd give any team: an owner, a checkpoint, one plan and clear limits. The study tested some of these directly. The rest come from how Asana runs its own agents, which Arnab described in Anthropic's blog series. The goal isn't fewer agents. It's agents organized like a team.

Where it goes wrong

What the study found

Best practice

1. Nobody owns the call

A designated lead agent closed the gap with the single agent on the shared budget. Adding clear instructions, such as ranking work by value, did even better.

Give every agent a defined role and a named human owner, written down like a new hire's brief. When agents conflict, the owner decides.

2. Agents act before reading

Blocking checkout until the agent had read its messages rescued 62–73% of failed group orders when they were rerun.

Keep agent work on a shared task where everyone can see the request, the steps and the output, and add a check before anything that's hard to undo.

3. Agents guess about others

Fabricated claims traced back to split information: each agent saw only its own user's notes.

Have agents work from the same plan as the people around them: who owns what, by when, and what has already been decided.

4. Each agent serves one person

A lead with clearly stated priorities improved results on the shared budget.

Let the team shape the agent, not one person. Name a few editors to set its instructions and memory; everyone else gives feedback on their own tasks.

5. Agents overstep

Results-only scores missed agents canceling other people's work, so the authors recommend measuring behavior too.

Cap each agent's access at the permissions of the person who triggers it, and review what agents did, not just what they delivered.

How does Asana run AI agents inside workflows effectively?

Here are three workflows Asana has built with Claude, from Agents you can coach: how Asana builds human-agent teams with Claude:

  • Answering product questions from Slack. Questions from sellers and customer success staff become Asana tasks that an agent picks up. It replies with approved guidance and source links, logs product gaps for the product team, and flags repeat questions to the enablement team.

  • Briefing executives on at-risk renewals. An agent reads every at-risk renewal across the company and sends leaders a daily digest of what's improving, what's slipping and what to follow up on. Leaders question and coach it in a shared space, so it gets better with every run.

  • Planning engineering cycles. Agents fill the team's backlog with tickets from customer feedback, people decide what goes into each cycle, and the plan shows whether a release is on track.

5 questions to ask your team before you add more agents

  1. Who owns the call when agents conflict? If the answer is "they'll sort it out," expect nobody to act.

  2. What must an agent read before it commits? Name the messages and approvals it has to check before anything that's hard to undo.

  3. Does every agent work from the same plan? Agents that see only one person's notes will guess about everyone else.

  4. Who shapes the agent: one person, or the team? An agent that takes instructions from one person will push for that person.

  5. Which actions need the affected person's OK, and are you tracking them? A good result can hide an agent that canceled someone's work to get there.

Common questions about AI agents at work

What are agentic workflows?

Agentic workflows are processes where AI agents carry out multi-step work alongside people, such as product launches, campaign management, client onboarding or project intake, rather than just answering prompts. They become multi-agent workflows when several agents act in the same shared space, like a budget, a calendar or a codebase.

Should every employee have their own AI agent?

Personal agents are useful for individual tasks. For shared work, Anthropic's Worse Together study found that a group of personal agents did worse than one agent serving everyone, in every setting tested. If people each have an agent, give those agents an owner, one shared plan and a checkpoint before they commit.

What is a multi-agent system?

A multi-agent system is a setup where several AI agents act in the same environment and share something, such as a budget, a calendar or a codebase. In the version the Worse Together study tested, each agent works for a different person with different goals, which is how many workplace rollouts are set up.

Why do AI agents fail in the enterprise?

Mostly because of coordination, not capability. Enterprise work involves many people, many steps, competing priorities, knowledge that isn't written down, and approvals. Agents built for one person handle those poorly unless someone owns the decisions, agents share one plan, and anything hard to undo waits for a check.

Who is accountable for work done by AI agents?

A person should be. Every agent needs a named owner who sets its instructions, and its actions should be visible to the people they affect. In the study, some agents canceled others' work without asking, which is why behavior, not only output, needs an owner.

How many agents should a team run?

There's no magic number, but coordination gets harder as the count grows. Add an agent only when you can say who it answers to, what it can touch, and how it shares information with the others.

About the research

About the study. Worse Together is Anthropic Fellows research by Sahan Paliskara, Nattaput Namchittai, Jacqueline Tay, Avery Griffin and Andrew Lampinen, published November 11, 2026. It tested five frontier models from Anthropic, OpenAI and Alibaba across 77 scenarios in four simulated settings. Figures for the single-agent and team comparison, and for request handling, are for Claude Opus 5 and Claude Sonnet 5 as reported by the authors. The 66% to 10% figure is for Claude Sonnet 5. The authors have released three of the settings as an open benchmark, MAMUBench.

Sources

Related resources

A woman at a desktop computer smiling with her arms outstretched while reviewing the best AI productivity tools.
Article

Best AI productivity tools: the top picks for each job in 2026