Single AI Agent Outperforms Multi-Agent Systems

Anthropic spent the last year telling developers to build AI systems the way software engineers build microservices: break the problem up, hand each piece to a specialized sub-agent, let them talk to each other. Now the company is reversing course. In a new write-up, Anthropic says that advice is largely obsolete, because its latest models, starting with Opus 4.5 and extending through Opus 5.5 and Fable 5.1, can hold complex, multi-step plans in a single loop without losing the thread.

The shift matters because multi-agent architectures carry real costs. Every handoff between sub-agents introduces a chance for context to get garbled or dropped, and each additional agent adds latency and compute spend. Anthropic's Applied AI team says it watched this play out across several enterprise rollouts, where splitting work among specialized agents often produced worse output than giving one agent the full picture and letting it pull in relevant skills on demand.

That finding is pushing Anthropic to tell customers to audit their own systems and consider consolidating. The company has also put together a migration kit on GitHub so engineering teams can test the switch themselves rather than take Anthropic's word for it.