Anthropic says one AI agent beats many subagents now that models can handle complexity alone

Anthropic is recommending that companies reassess multi-agent architectures, arguing that a single primary agent which pulls in context as needed frequently beats a fleet of specialized sub-agents on quality, cost and latency. The guidance draws on the company's Applied AI team's hands-on work across multiple enterprise deployments. Anthropic has also published a migration kit on GitHub that walks teams through identifying existing patterns, evaluating alternatives and verifying which approach works best.
The technical argument rests on newer models. Anthropic says models beginning with Opus 4.5, along with Opus 5.5 and Fable 5.1, can carry complex plans directly inside one loop rather than delegating steps to separate agents. In a single architecture comparison cited by the company, the single-agent setup ran about twice as fast as the multi-agent equivalent. Anthropic's Applied AI team found that a single agent using so-called skills often outperformed both monolithic prompt designs and sub-agent setups on quality.
The failure modes Anthropic describes are familiar to anyone who has wired agents together. Context is lost at every boundary, so in a workflow that moves from feature writing to test writing to code review, each handoff sheds information the next step needed. Those losses show up as higher latency and rising cost, because each hop means another model call with its own prompt overhead. In the single-agent pattern Anthropic illustrates, one agent writes the feature and its tests together since it already holds the context, while a separate verifier only runs the tests and reports results, a much narrower role than a full specialist agent.
The organizational story is harder to fix than the technical one. Anthropic describes a pattern where intent detection, a fixed processing pipeline, user segmentation, separate recommendation branches and an orchestrator on top get built out even though the model can now handle much of that work itself for the use case at hand. The company calls the result a Conway's Law effect for agents: the architecture comes to mirror the org chart, with each team shipping its own sub-agent until hundreds of overlapping agents accumulate without shared standards. The follow-on is what Anthropic terms a Frankenprompt, in which every team appends rules to the system prompt and the rules contradict one another. According to the company, a clean system prompt, four or five lean data tools and fixed output guardrails cover much of a given use case.
Anthropic points to its own product as evidence. It cut the Claude Code system prompt for newer models by more than 80 percent, partly because skills and tools now load only when needed rather than being declared upfront. Newer models also narrate their own progress and check their own work, which the company says can make some of the extra control layers redundant.
One enterprise example involved a company whose customer-facing agent contained a dozen specialized sub-agents. That structure made a consistent brand experience difficult to maintain and hurt usability through high latency, according to Anthropic. This is the practical cost of the org-chart pattern: every team's local optimization becomes a global slowdown for the end user.
Sub-agents are not written off. Anthropic says they still make sense in three cases: when a single agent's context window shows measurable performance degradation, when multiple tasks must run in parallel across separate context windows, and when a single agent struggles with tool selection.
The migration itself carries risk. Anthropic warns against merging blindly and describes a test with a support agent where folding in a sub-agent did reduce latency but broke a previously working process, because the primary agent began answering on its own instead of consulting the right specialist. In practice, that suggests teams should keep behavior evaluations around routing and escalation, not just latency and cost, before deleting an agent.
The company's recommendations for keeping a consolidated system maintainable are concrete. Rather than giving every team its own agent, package workflows into skills and load tools on demand through a tool search. A small central team should own the system prompt and overall quality. Anthropic notes that the model's tool selection tends to degrade beyond roughly 30 to 50 tool definitions, which puts a hard ceiling on how much a shared tool surface can grow before accuracy suffers; designated owners on each team should maintain skill libraries so they do not balloon. 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, or embedding that material as reference files inside a skill. It advises against fetching context through code and injecting it straight into prompts.
The findings line up with outside work. A study by Vals found agent teams can be faster than single agents but do not deliver higher quality, and that scaling from ten to 30 or 100 agents produced no quality gains while significantly raising costs. OpenAI developer Eric Provencher has likewise questioned agent swarms, calling them almost certainly wasted money. For developers, the takeaway is that parallelism buys wall-clock time, not correctness, and the bill scales with agent count.
Why it matters: Teams that shipped a sub-agent per team, per domain or per pipeline stage now have a vendor-endorsed case for consolidating, and the migration kit plus the 30-to-50 tool ceiling give them something concrete to test against. The shift moves design effort from orchestration graphs toward skills, tool search and a single well-owned prompt, and it means evaluations should track routing and escalation behavior so a latency win does not quietly break a working process. It is reasonable to infer that the pattern will spread unevenly, since workloads with genuinely parallel, context-heavy tasks still fit the multi-agent shape Anthropic itself carves out.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Anthropic says companies should evaluate whether they can consolidate sub-agents into a single agent that loads context on demand. In one architecture comparison, the single-agent setup ran about twice as fast. Anthropic has published a migration kit on GitHub to help teams make the transition. The article Anthropic says one AI agent beats many subagents now that models can handle complexity alone appeared first on The Decoder .