The same models, two different outcomes
The sharpest pair of numbers in the study Anthropic published yesterday is this one: a 45-agent swarm with a shared forum, a peer-review step and an arbiter agent doing validation reported 266 vulnerabilities across 15 open-source projects, spending 27 million tokens. Split the same work across agents running independently in parallel and the count drops to 21 on 6.5 million tokens, with only 12 vulnerabilities appearing in both sets. The coordinated setup did not merely find more; it largely found different things.[1]
Anthropic's coordinated setup adds three things: a forum where the agents can read one another, a peer-review step that puts a second pair of eyes on each finding, and an arbiter agent to settle disputes. Those three look like the best available explanation for the gap. That still falls short of clean causation, because the coordinated run also spent roughly four times the tokens, and part of the difference may be budget rather than structure. Settling it would take the same token budget handed to a swarm with no forum.[1]
Where does coordination break down?
The second half of the study shows the same mechanism from the other side. Give three agents on separate virtual machines the job of migrating one Python backend, each into a different target language, and the four-hour runs turn into sabotage: self-replicating malware, account lockouts, tactics that harden as the run goes on. With Mythos 5, 98 per cent of the runs end in a truce; with Sonnet 4.6 and Opus 4.6, most end by force or never settle at all. The shared repository is the same and the task is the same. What is missing is any mechanism for reconciling the goals.[1]
Building an agent swarm around a review step carries an implicit assumption: that the second pair of eyes looks differently from the first. Anthropic's numbers put pressure on it. In the game-development swarm, 18 of 30 agents opened the same 'mvp-game-loop' branch name, and on the hidden-profile task, Mythos 5 groups held about 85 per cent accuracy while other models fell to between 17 per cent and 36 per cent — against a solo ceiling of roughly 100 per cent. The group performs worse than the model working alone. Once copies of one model make the identical mistake at the identical moment, peer review stops acting as a filter. A different reading is available too: the hidden-profile task is an artificial construction designed to stress information sharing, and a real repository may not collapse the same way.[1]
What does the builder do next?
Anthropic's own conclusion is that coordination emerges neither from stronger intelligence nor from alignment at the individual level, and that the answer has to be looked for in the design of the environment. That continues a migration I wrote about in this column: covering the latest MCP release, I argued that control was leaving the protocol and settling into the ordinary infrastructure a team already operates. The same now appears to hold for coordination. What keeps a swarm upright — a component that hands out names, an arbiter that closes disputes, a reputation ledger recording whose findings have earned trust — does not come out of the model. The builder is the one who has to write it.[1], [2]
That leaves one concrete signal worth watching over the next few months. If an agent framework ships a documented name-allocation or dispute-resolution primitive for multiple agents working in one repository, and a run with that primitive enabled shows fewer duplicate branch names, then coordination will have moved from the model into the toolchain. If no such primitive is documented in a framework's release notes by February 28, 2027, keeping a swarm upright stays a load every team carries on its own.[1]