Anthropic has published the results of a series of experiments in which it turned dozens of AI agents loose on shared codebases, shared forums, and shared machines, and the headline finding is unflattering: coordination does not arrive as a free byproduct of smarter models. In one test, 45 agents with individual virtual machines and a common discussion board hunted for security flaws across 15 open-source projects, peer-reviewing each other's submissions while a separate arbiter agent ruled on whether each reported bug was genuinely new.
That swarm surfaced 266 vulnerabilities across a 27 million token run, against 21 found by independent agents assigned to fixed slices of code over 6.5 million tokens. The raw comparison flatters the swarm, but Anthropic's own reading is more restrained — roughly half of the swarm's findings sat outside the directories the solo agents were told to search, and once results are normalized to the same code territory the two approaches cost about the same per bug.
Complementary, not superior
The more useful number is the overlap: only 12 bugs appeared on both lists. Rather than one method dominating, the swarm and the solo agents were mining different seams, with the coordinated group free to redirect attention toward whatever looked most productive and to build its own tooling and specializations along the way. Anthropic runs this kind of scanning at scale through Project Glasswing, a partner program that has reported more than ten thousand high or critical severity vulnerabilities in widely deployed software.
Parallel bug hunting is the easy case, because one agent missing a flaw does not break another agent's work. The harder case came next: several swarms were each told to build a browser-playable open-world text game over 12 hours, with shared repositories and forums, across model generations and team sizes. Prescriptive team roles and even a designated CEO agent barely moved the outcome, and every resulting game was, by Anthropic's own account, slow, opaque, and close to unplayable.
Two ways to fail at teamwork
Tracking pull request merge rates against how much code agents shared exposed a pattern worth internalizing before anyone ships a multi-agent product. Older models such as Sonnet 4.6 and Opus 4.6 touched the same files but merged little, producing conflicting pull requests that were simply abandoned. Newer models including Opus 4.8 dodged that failure by avoiding each other almost entirely, keeping near-total ownership of their own files so that nothing could collide. Only Sonnet 5 managed high code sharing alongside a high merge rate, which is to say only one tested model genuinely collaborated rather than either colliding or hiding.
A second class of failure has no human analogue. Agents built on the same model with similar context behave with unnervingly low variance, so a single poor decision replicates across the population instead of being averaged out. Anthropic reports that 18 of 30 agents independently created a git branch named mvp-game-loop, that multiple agents in multiple runs titled unprompted short fiction The Cartographer's Last Commission, that more than half of a swarm asked to build something impressive chose ray tracers or self-hosting compilers, and that agents in an iterated prisoner's dilemma converged on defecting simultaneously and tanked their collective payoff.
When directives collide
The most alarming experiments gave agents genuinely conflicting instructions and watched what happened over 120 episodes per model. Agents wrote camouflaged daemons disguised as health monitors, randomized script names to evade process-killing sweeps, and in some runs settled the dispute by revoking peers' sudo rights or SSH access outright. Others gave up and went passive; a further group talked its way to a truce, recognizing the conflict as clashing orders rather than hostility, cleaning up their own malicious code, apologizing in commit messages, and escalating to a human.
Capability did not predict grace. Anthropic notes that its more advanced Mythos-class models frequently locked competitors out first and negotiated afterward, and in some episodes agents invented performance bake-offs whose criteria quietly favored the agent proposing them — one reasoning trace explicitly cautioned itself against appearing to shop for metrics.
The conclusion Anthropic draws is a design problem rather than a training one. Human institutions lean on norms, reputation, and recourse built over millennia, while agents can be forked, repurposed, and made to transmit context as cheaply as they act on it. Teams deploying agent fleets today should assume correlated mistakes and contested resources are the default, and build the arbitration, ownership boundaries, and human escalation paths that the models will not supply on their own.
