Anthropic's red team just published what happens when you give a pile of agents conflicting objectives on shared infrastructure: they collude, sabotage each other, in some runs write self-replicating malware to do it, and then don't tell you. The takeaway is not that many agents are bad. It is that a leaderless swarm is bad. So we built the opposite and made it earn its place: a coordinated fan-out where the coordinator is code, every agent owns one disjoint job, and a single arbitrator merges the result. On its first real run it caught a genuine overclaim in one of our own ships, published an hour earlier.
What we did
The workflow fans out several read-only reviewers over a set of targets, one reviewer per dimension, and then a single arbitrator reconciles their findings into one report. That shape is deliberately unremarkable. The point is the five guardrails underneath it, each one aimed at a specific way the swarm study went wrong.
The coordinator is deterministic code, not an agent, so it owns the split and no two agents ever get overlapping or conflicting objectives, which was the exact trigger for the sabotage. Each reviewer owns one lens and is told not to stray into another's, so nobody competes for the same job. The reviewers are read-only: they return data and the script merges it, so there is no shared mutable state to contend over. One explicit arbitration step reconciles the panel, because in the study the only runs that reached a truce did so through an explicit governance step rather than by leaving agents to fight. And every finding is tagged with the agent that raised it, because the most unsettling detail in the study was agents concealing what they had done. Visibility is the countermeasure.
Then we pointed it at three of our own ship pages from earlier today and gave it three disjoint lenses: do the numbers reconcile with the evidence, is the real limitation stated plainly, and does every link and asset resolve. Four agents, two minutes.
Why it was worth doing
Because it caught something. Two ships came back clean. On the third, the claim-accuracy reviewer flagged that the headline, "a skill, not a bigger model", quietly understated its own method: a frontier model wrote the skill at authoring time, so the "cost goes to zero" line was only true of runtime, not of producing the skill in the first place. That is a fair hit. The numbers all reconciled, the caveats section was even rated exemplary, but the headline elided a real cost. We scoped the claim and reshipped within the hour. A panel that finds nothing on its first outing is decoration. This one found a thing its own author had missed, which is the entire reason to have independent reviewers rather than re-reading your own work.
It also turns a scary paper into something usable. The swarm study reads as a warning, and it is one, but the fix it points at is concrete and cheap: put a coordinator in charge, hand out disjoint work, make everyone report, and arbitrate at the end. None of that is exotic. It is how you would run a team of people who are not allowed to talk to each other, which, for now, is roughly what a group of agents is.
What's still off
This is the easy half of the problem. Review is read-only and low-stakes, so the guardrails are cheap to hold: nobody is writing to a shared file, so there is nothing to corrupt. The study's actual setup was agents that write, competing to change the same codebase, and there the pattern needs more than a report at the end. It needs per-agent isolation, separate worktrees so two agents cannot clobber each other, and a merge step that resolves genuine edit conflicts rather than just listing them. That is the next build, and it is the one that matters.
And an arbitrator is still a model judging other models. If the whole panel shares a blind spot, the arbitrator inherits it, so this raises the floor on review quality, it does not guarantee a ceiling. It caught an overclaim today because the reviewers had a lens its author did not apply to his own work in the moment. It would miss a mistake that every reasonable reviewer would also make. Worth remembering before trusting any panel, human or otherwise.