Anthropic is advising companies to test whether one primary AI agent can replace a collection of specialized subagents. Its applied AI team says newer models can manage longer plans and load domain instructions as needed, reducing the coordination overhead of a multi-agent design.
In deployments reviewed by the company, handoffs between agents lost context and added latency and cost. One customer had about a dozen specialists inside a customer-facing system, making tone inconsistent and responses slow. In one architecture comparison, Anthropic says a single agent completed the workflow about twice as fast.
The proposed alternative uses reusable “skills”: packages of instructions and context that the main agent loads only when relevant. The same agent can then write a software feature and its tests, for example, without passing an abbreviated summary to another model. Anthropic has published a migration kit to help teams compare the approaches.
The recommendation is not a universal rule. Separate agents can still provide isolation, parallel work, different permissions or independent checks. A single agent also creates one larger failure domain and may accumulate too much context if skills are poorly designed. Teams should measure task success, latency, token cost and information loss on representative workflows before changing architecture. The practical lesson is to earn each additional agent with evidence rather than assuming that a more elaborate organization is automatically more capable.