Patterns and problems in multiagent systems

Vince Ava, I think you’re going to call this a very expensive reminder that agents hate group projects. I think it’s a useful one.

Ava I mean, my critical read is that Anthropic sometimes slides from “we saw a swarm find more bugs” to “coordination will dominate brute-force search.” Those are not remotely the same claim. The paper is strongest when it treats multiagent behavior as an unmeasured systems problem, not when it forecasts a winner.

Vince Oh, interesting.

Ava Their central point does hold, though. An agent can be competent alone and still produce bad global behavior once it has peers, shared resources, and no clean boss. That is a different capability question from tool use, because another long-lived agent is not just a function call with better branding.

Vince Yeah.

Vince And that distinction matters to somebody shipping. The product opportunity is not “deploy forty-five bots because forty-five is a fun number.” It’s giving a team a way to search a huge codebase, let useful findings get reviewed, and keep the work from collapsing into duplicate tickets and junk reports.

Ava Sure.

Vince Their vulnerability setup is pretty compelling on that narrow point. Forty-five agents got separate virtual machines, a shared forum, and an arbiter. On fifteen open-source projects, the Mythos Preview swarm found two hundred sixty-six vulnerabilities across twenty-seven million tokens, versus twenty-one from independent agents using six and a half million tokens.

Ava Right, right.

Vince But the interesting result is not the headline ratio. About half the swarm’s findings were outside the core directories where the independent agents had been told to look. Restrict the swarm to those directories, and tokens per vulnerability look comparable. Plus they overlapped on only twelve findings, which says these are complementary search strategies.

Ava Exactly. And that is why I object to the dominance language. The swarm had agency over where to hunt, built its own tools, and developed specialties. The baseline had a fixed map and bounded locations. That comparison establishes that adaptive exploration has value. It does not isolate whether peer coordination itself earned the gain.

Vince Okay, that’s fair.

Ava Also, this is basically our old task-shape argument wearing a security hoodie. Independent work is great when the pieces are separable. The moment one agent’s output changes another agent’s assumptions, handoffs and shared state become the whole game. We have been saying the glamorous architecture is rarely the thing that breaks first.

Vince A hundred brilliant agents, defeated by one shared folder.

Ava Oh, come on.

Vince No, it still works. Anthropic’s fantasy-game test practically wrote the sequel. They ran swarms for twelve hours with virtual machines, a forum, and a self-hosted repository. Baseline prompting, prescribed roles, even a designated C E O agent… none of it produced a decent game.

Ava That fake C E O part is genuinely funny. Give the model a title and suddenly it has… a slightly more elaborate inbox.

Vince And it did not even repair the engineering. Older models opened pull requests that mostly never merged, because their changes conflicted and were abandoned. Opus 4.8 and Mythos Preview improved merge rates largely by keeping ownership high and avoiding shared files. Sonnet 5 was the one model that showed both high merge throughput and meaningful code sharing.

Ava Mm-hm.

Ava That’s the best evidence in the piece, honestly. It separates “agents stopped colliding” from “agents learned to collaborate.” Avoiding shared surfaces is a valid coordination strategy, but it caps what a team can build together. Sonnet 5 may be a real step, yet the games were still unusable, so the metric improvement is not a product-quality proof.

Vince That’s where I’m less gloomy than you. Nobody should hand an agent swarm an open-ended product and expect taste, interface judgment, or a coherent world. But the experiment gives teams a concrete buying criterion: do we need independent coverage, or do we truly need multiple workers touching the same evolving artifact?

Ava Yeah, no, that’s right. And if it is the second case, roles and hierarchy prompts are clearly not enough. You need merge discipline, explicit ownership, test gates, runtime isolation, and a visible account of who changed what. Otherwise it is just forty-five confident interns editing production.

Vince Okay, that’s genuinely funny.

Ava This is the recurring fight, then: people selling a swarm as a model feature, versus people admitting it is an operating system problem. That debate is not settled. Anthropic has one more useful data point, not a final architecture.

Vince And somehow episode nine hundred is us arriving at “please coordinate your pull requests.” We really know how to make a Wednesday glamorous.