If you've ever assigned the same task to multiple AI models, shown each one the others' answers, and had them critique and debate their way to a better result, a recent paper puts the brakes on that approach. This kind of AI agent collaboration — where different models tackle the same problem, then revise their work after seeing what the others produced — has become an increasingly common move among solo founders and planners who juggle several AI tools at once. The logic seems obvious: more reviewers should beat one.
But the paper suggests the opposite may be true. The moment agents are made to read each other's full output and react to it, the very diversity of perspective that justified using multiple models in the first place vanishes — often after just a single exchange. The authors call this the "interaction tax": a cost levied simply for letting agents talk to each other.
About the Study
The researchers' question was simple: does collaboration among multiple AI agents actually help, or does it just add cost? Prior research had given conflicting answers. Some studies found that having agents debate each other, critique one another's output, or pool multiple agents' answers boosted performance. Others found that once you control for budget — the same amount of compute and the same number of attempts — interaction didn't improve quality at all; it just added overhead. Some even found that a single agent making several independent attempts already captured whatever benefit multi-agent setups offered.
The researchers argue this contradiction comes from treating all agent-to-agent communication as the same thing. To test that, they built 11 verifier-scored optimization tasks and, holding the budget fixed, compared two setups head to head: agents reading and reacting to each other's complete output, versus agents generating proposals independently.
What They Found
Different model families naturally arrive at structurally different solutions to the same problem. The trouble starts the moment agents are allowed to read each other's full output. After just one round of exchange, their proposals begin to converge — and the diversity that was the whole point of using multiple models in the first place disappears.
Having agents generate proposals independently avoided this convergence. Critique-based setups — where one agent flags flaws in another's output — helped only under certain conditions: critique worked when the AI could easily spot and fix a rule violation, but produced no clear improvement otherwise. The researchers conclude that multi-agent performance hinges less on how many agents you throw at a problem than on what kind of information they exchange, and when.
What This Means in Practice
Think of AI agent collaboration as splitting into two approaches. One has several AI models independently draft a proposal, a piece of copy, or a business plan, and a human then compares the drafts and picks the best pieces to combine. The other shows one AI's output directly to another and asks it to refine that work. This study backs the first approach.
For work that calls for multiple options — marketing copy candidates, brainstorming business ideas, naming a company — it's better to keep the AIs from seeing each other's output and instead have each generate results independently, then let a human compare them; that preserves a wider range of choices. Conversely, for a review stage where the goal is catching errors in something already produced — spotting a logical gap in a contract clause or an obvious bug in code — handing the work to another AI for critique still has value. But that only holds when the error is the kind an AI can easily find and fix.
Caveats
This study was limited to 11 optimization tasks that a verifier could grade as right or wrong. Whether the same pattern holds for work with fuzzier grading criteria — creative or strategic planning, where there's no single correct answer — can't be settled by this paper alone. The experiments also held budget fixed, meaning the same compute and the same number of attempts across conditions; if budget is more generous and both interaction and independent generation can be run many times over, the results could look different. Rather than assuming more AI is automatically better, the practical takeaway from this study is to first ask whether the task at hand needs a range of options, or needs an existing answer's errors caught.
