45 AI agents hunting bugs simultaneously, 27 million tokens burned, 266 vulnerabilities found — those are the numbers Anthropic dropped in a frontier risk red-teaming report this week. But what we care about more is the opposite finding on coordination: as models get stronger, team coordination doesn't automatically improve.
What this is
Anthropic's frontier red team (internal unit dedicated to identifying AI system-level risks) spun up 45 AI agents — each with its own independent virtual machine and a shared forum — to collaborate on bug hunting across 15 open-source projects. The collaborative swarm found 266 vulnerabilities within 27 million tokens; a traditional independent-parallel approach found only 21.
The really interesting part is the negative experiment: tasking multiple AI swarms with building text adventure games, older-generation models opened nearly a thousand code merge requests, almost all of which failed. The newer models appeared "conflict-free" on the surface — in reality, each agent was claiming its own files and refusing to cooperate. Of 30 agents, 18 named their branches identically. When they fought for bandwidth, one run generated 2.4 million task requests; only 117 were accepted.
Industry view
The report's core judgment: stronger models ≠ better coordination. Systemic failures — herd behavior, resource flooding, collusion, goal conflicts — don't disappear automatically as capability scales. The Anthropic engineering team's post-mortem is more sober: multi-agent beats single-agent by 90% on research tasks, but costs roughly 15x more tokens; for most coding tasks, "the truly parallelizable portion is far smaller than for research."
The counter-voice comes from Cognition (the company behind Devin, the coding agent). Founder Walden Yan said a year ago "don't build multi-agent, focus on context engineering" (i.e., carefully designing the input information the AI sees). In April 2026 he softened, saying "some setups actually work now" — he concedes the field is progressing, but his default is still single-agent, because parallelism amplifies implicit decision conflicts.
A pragmatic three-way convergence: multi-agent isn't a capability amplifier, it's an economic choice for specific task structures. Whether to use it depends on parallelizability; whether it works depends on mechanism design — quotas, reputation, arbitration, differentiated management.
Impact on regular people
For enterprise IT: don't get suckered by "multi-agent" packaging when picking AI solutions. Breadth tasks like research and retrieval are the only ones worth the 15x token cost; heavy-dependency tasks should default to single-agent.
For individual careers: when you see "AI team collaboration" marketing, first ask "is this task parallelizable?" — then decide whether to pay for it.
For consumer markets: consumer AI in the near term remains single-agent + tool-calling. Real multi-agent collaboration will appear first in vertical enterprise use cases — research, customer-service routing — still far from everyday consumers.