Anthropic is urging companies to ask whether they can replace a fleet of specialized subagents with a single primary agent. According to the company, an agent that loads context only when needed often delivers better results in quality, cost and response time. In a comparison cited by Anthropic, the single-agent architecture ran about twice as fast as the multi-agent setup.
The recommendation comes from hands-on work by Anthropic's Applied AI team, which found agents using so-called skills often outperformed both monolithic prompts and subagent setups on quality. One company had a dozen specialized subagents inside its customer-facing agent, producing an inconsistent brand experience and high latency.
The handoff problem
In multi-agent workflows, context is lost at every handoff between writing features, writing tests and code review. In a single-agent workflow, one agent writes both feature and tests because it already holds the context, while a separate verifier only runs the tests and reports results.
Anthropic also describes an organizational effect it calls "Conway's Law for agents": the architecture ends up copying the org chart, with each team shipping its own subagent until hundreds of overlapping agents pile up without shared standards. The result is a "Frankenprompt" full of conflicting rules. According to the company, a lean system prompt, four or five data tools and fixed output guardrails cover much of most use cases.
When subagents still make sense
Anthropic says subagents still make sense in three cases: when a single agent's context window shows measurable performance degradation, when several tasks must run in parallel with separate contexts, or when a single agent struggles to pick the right tool.
For migration, Anthropic has published a kit on GitHub that guides teams through identifying existing patterns, evaluating alternatives and verifying the best approach. It warns against merging blindly: in one test with a support agent, folding in a subagent cut latency but broke a working process, because the primary agent began answering on its own instead of consulting the right specialist.
Keeping skill libraries in check
Instead of one agent per team, Anthropic suggests packaging workflows into skills and loading tools on demand through a tool search. A small central team should own the system prompt and overall quality. According to the company, a model's tool selection tends to degrade beyond 30 to 50 tool definitions, so each team should have an owner maintaining its skill library.
For reference material such as guidelines, product specs or metric definitions, Anthropic recommends knowledge bases the agent queries only when needed through a search tool. Such information can also be embedded as reference files inside a skill.
The findings are consistent with recent research, cited by The Decoder, which found agent teams can be faster than single agents without delivering higher quality: scaling from ten to thirty or a hundred agents brought no quality gains but much higher costs.
No comments yet. Be the first!