Anthropic finds a single AI agent often outperforms a fleet of sub-agents

Anthropic reports that a single agent loading context on demand frequently beats multi-agent setups in quality, cost, and speed. The company is releasing a migration kit to help teams evaluate and make the switch.

Experience across multiple enterprise deployments shows that a primary agent using skills often beats both monolithic prompt designs and sub-agent setups. One company, for example, had a dozen specialized sub-agents inside its customer-facing agent. This made it hard to keep a consistent brand experience and hurt usability with high latency.

Newer models starting with Opus 4.5, along with Opus 5.5 and Fable 5.1, can manage complex plans directly in a single loop. This reduces costs and wait times. In one comparison, Anthropic says the single-agent architecture ran about twice as fast as the multi-agent setup.

The model can now do much of the orchestration work itself

Many agent designs follow a familiar pattern with intent detection, a fixed processing pipeline, user segmentation, separate recommendation branches, and an orchestrator sitting on top. Much of that is work the model can now handle on its own for the respective use case.

Anthropic calls it “Conway’s Law” for agents. The architecture becomes a copy of the org chart, with each team shipping its own sub-agent until hundreds of overlapping agents pile up with no shared standards. What follows is what Anthropic calls a “Frankenprompt,” where every team adds rules to the system prompt that end up contradicting each other. A clean system prompt, four or five lean data tools, and fixed output guardrails cover much of the respective use case, according to the company.

As one example, Anthropic cut the Claude Code system prompt for newer models by over 80 percent, partly because skills and tools now load only when needed. Newer models also comment on their own progress and check their own work, which can make some of the extra control layers unnecessary.

Anthropic says sub-agents still make sense when a single agent’s context window shows measurable performance degradation, when multiple tasks need to run in parallel with separate context windows, or when a single agent struggles with tool selection.

For a step-by-step migration, Anthropic has published a migration kit on GitHub that walks teams through a workflow to identify existing patterns, evaluate alternatives, and verify the best approach. Anthropic warns against merging blindly, since in one test with a support agent, folding in a sub-agent did reduce latency but broke a previously working process. The primary agent started answering on its own instead of consulting the right specialist.

Teams should keep skill libraries from sprawling

Instead of giving every team its own agent, Anthropic recommends 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.

Separately, Anthropic notes that the model’s tool selection tends to degrade beyond 30 to 50 tool definitions, so designated owners on each team should maintain the skill libraries to keep them from ballooning.

For reference material like guidelines, product specs, or metric definitions, Anthropic recommends knowledge bases that the agent queries only when needed through a search tool. That kind of information can also be embedded as reference files inside a skill. Anthropic advises against fetching context through code and injecting it directly into prompts.

Anthropic’s findings are consistent with recent research. A study by Vals found that agent teams can be faster than single agents but do not deliver higher quality. Scaling from ten to 30 or 100 agents brought no quality gains but significantly higher costs, according to the study. OpenAI developer Eric Provencher has also questioned the value of agent swarms, calling them “almost certainly” wasted money.

What it means

Developers building customer support or internal tools should test whether a single agent can handle their specific workflow before splitting tasks across multiple bots. The new models handle planning and self-checking well enough that extra layers of control often add cost without adding value.