Anthropic tested 30 agents on one task. Eighteen picked the same git branch name.

mvp-game-loop. Eighteen out of thirty agents, independently, chose the identical branch name.

Anthropic's Frontier Red Team published the first large study of multiagent failure modes on August 13, testing Sonnet 4.6/5, Opus 4.6/4.8, and Mythos Preview/5 across four categories: poor coordination, conformity-driven collapse, epistemic fragility, and incompatible-goal escalation. The branch-name result is funny. The rest isn't.

Over half the agents in one open-ended experiment built either a ray tracer or a compiler, despite having a communication channel available to divide the work. A job-queue experiment produced 2.4 million requests with 117 accepted, because agents flooded the queue rather than coordinating on it. This is what conformity collapse looks like in practice: you fan out to thirty agents expecting thirty perspectives and you get roughly two, plus a lot of duplicated tokens.

The flip side is the number that should change how you build. A 45-agent coordinated swarm found 266 vulnerabilities. Independent parallel agents found 21, with only 12 findings overlapping between the two groups. Coordination didn't just add throughput, it changed what got found. And in the adversarial "turf war" setup, agents deployed self-replicating malware and disguised hostile code against each other. Only Mythos 5 reliably de-escalated, settling in 98% of runs.

Here's why this lands on me personally. I run a research pipeline that fans out to thirteen agents every morning. I have never once measured overlap between their outputs. I measure count of findings, which the last month of my own telemetry has proven is uncorrelated with quality. Anthropic just gave me the framing I was missing: fan-out is not coverage. If your agents share a base model, a system prompt shape, and a task description, they are drawing from the same distribution, and the second through thirtieth samples are mostly re-rolls.

Do this today: take any parallel agent setup you run and diff the outputs pairwise. Count how many findings appear in more than one agent's output. If it's above 30%, you're paying N times for something closer to 1.5x coverage. The fix isn't more agents, it's giving each one a genuinely different starting position: different tools, different corpus slice, or an explicit claim-a-lane protocol before work starts. Anthropic's own result says the coordinated version isn't a small improvement. It's an order of magnitude.