Anthropic says one AI agent often beats many subagents, releases migration kit

Anthropic is recommending that companies evaluate whether they can consolidate sub-agents. Its argument: a single primary agent that loads context as needed often delivers better results than a fleet of specialized ones. Anthropic says a single agent with on-demand context often outperforms multi-agent setups in quality, cost and latency. The company has published a migration kit on GitHub to help teams evaluate and make the switch. The account below follows The Decoder's report.

The reasoning is that newer models, starting with Opus 4.5 and including Opus 5.5 and Fable 5.1, can manage complex plans directly in a single loop, which can also cut costs and wait times. In one comparison, Anthropic says the single-agent architecture ran about twice as fast as the multi-agent setup. The recommendation comes from hands-on experience across multiple enterprise deployments, where Anthropic's Applied AI team found that a single agent using so-called skills frequently beat both monolithic prompt designs and sub-agent setups on quality. One company, for example, had a dozen specialized sub-agents inside its customer-facing agent, which made it hard to keep a consistent brand experience and hurt usability through high latency.

According to Anthropic, many agent designs follow a familiar pattern: intent detection, a fixed processing pipeline, user segmentation, separate recommendation branches, and an orchestrator on top. Much of that is work the model can now handle itself. Anthropic calls the result "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. Next comes what Anthropic calls a "Frankenprompt", where every team adds rules to the system prompt and the rules end up contradicting each other. In Anthropic's view, a clean system prompt, four or five lean data tools and fixed output guardrails cover much of the use case. As one example, it 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 extra control layers unnecessary.

Sub-agents 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 context windows, or when a single agent struggles with tool selection.

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

On organisation, Anthropic recommends packaging workflows into skills and loading tools on demand through a tool search, instead of giving every team its own agent. A small central team should own the system prompt and overall quality. Because the model's tool selection tends to degrade beyond 30 to 50 tool definitions, designated owners on each team should maintain skill libraries so they don't balloon. For reference material such as guidelines, product specs or metric definitions, Anthropic recommends knowledge bases the agent queries through a search tool only when needed, or reference files embedded inside a skill. It advises against fetching context through code and injecting it directly into prompts.

The Decoder says the findings are consistent with recent research. A study by Vals found that agent teams can be faster than single agents but don't deliver higher quality; scaling from ten to 30 or 100 agents brought no quality gains but significantly higher costs. OpenAI developer Eric Provencher has also questioned the value of agent swarms, calling them "almost certainly" wasted money.

Key facts

Anthropic says a single agent with on-demand context often outperforms multi-agent setups in quality, cost and latency, and urges companies to evaluate consolidating sub-agents.
It has published a migration kit on GitHub that walks teams through identifying existing patterns, evaluating alternatives and verifying the best approach.
In one comparison the single-agent architecture ran about twice as fast; Anthropic cut the Claude Code system prompt for newer models by over 80 percent.
Sub-agents still make sense for context-window degradation, parallel tasks needing separate context windows, or tool-selection trouble; tool selection tends to degrade beyond 30 to 50 tool definitions.
Blind merging is risky: in one support-agent test, folding in a sub-agent cut latency but broke a working process.

Why it matters

Multi-agent designs with an orchestrator and many specialist sub-agents have been a common way to build agent products. Anthropic is now telling teams that, with newer models (Opus 4.5 onward, plus Opus 5.5 and Fable 5.1), much of that structure is unnecessary and can cost more in latency, spend and information lost at each handoff. Its own example: it cut the Claude Code system prompt for newer models by over 80 percent.

Who it affects

Teams running enterprise agents, especially those where each department has built its own sub-agent. Anthropic's example is a company with a dozen sub-agents inside one customer-facing agent. It also affects whoever owns the system prompt: Anthropic suggests a small central team for that, with designated owners keeping each team's skill library from growing too large.

How to use it

Anthropic's kit on GitHub walks through a workflow to identify existing patterns, evaluate alternatives and verify the best approach. The advice in the report: package workflows as skills, load tools on demand through a tool search, keep a clean system prompt with four or five lean data tools and fixed output guardrails, and put reference material in knowledge bases queried by a search tool or in reference files inside a skill. Keep sub-agents where context degrades measurably, where tasks run in parallel with separate context windows, or where tool selection suffers.

How solid is it

The claims come from Anthropic's Applied AI team and its experience across multiple enterprise deployments, relayed by The Decoder rather than read in an original Anthropic document. The report gives no cost or quality figures for the comparisons; the roughly twofold speed gain comes from one comparison. The Decoder points to the Vals study and Eric Provencher's comments as consistent with Anthropic's view, but gives no further detail on the study.

Risks and caveats

Anthropic itself warns against merging blindly. In a support-agent test, folding in a sub-agent reduced latency but broke a previously working process, because the primary agent answered on its own instead of consulting the right specialist. Sub-agents remain the better choice in three cases: context-window degradation, parallel work with separate context windows, and tool-selection problems. Anthropic's recommendation is to evaluate consolidation, not to apply it everywhere.

“Anthropic found that a single agent with on-demand context often outperforms multi-agent setups in quality, cost, and latency.”

— The Decoder, summarising Anthropic's findings