Anthropic pushes to reduce agent swarms in favor of a single one
Anthropic recommends teams to review their AI agent architectures. According to the company, a single agent that loads context on demand can be cheaper, faster, and better at retaining information than multiple sub-agents.
Anthropic believes that a single AI agent can now outperform a sub-agent architecture in many enterprise use cases. The company recommends teams assess whether they can consolidate their sub-agents around a single main agent that loads context on demand. Based on its experience, this approach often reduces costs, latency, and information loss at handoff points. This topic matters because it directly affects how companies design their assistants, workflows, and AI budgets.
Quand moins d’agents fait plus de sens
Anthropic reports having observed, in several deployments, that a single agent with tools and "skills" loadable on demand can outperform multi-agent systems in quality, cost, and response time. In a comparison cited by the company, the single-agent architecture ran about twice as fast as the multi-agent version. The reasoning is simple: at each transfer between sub-agents, some context is lost, and coordination becomes heavier. As a result, an agent meant to streamline work sometimes ends up resembling a meeting room with too many participants. It’s amusing, but the cloud bill is not.
The recommendation mainly targets organizations that have stacked specialized functions for intent, segmentation, recommendation, or orchestration. Anthropic explains that recent models, starting with Opus 4.5, then Opus 5.5 and Fable 5.1, can handle complex plans alone in a single loop. In this context, a leaner architecture — light system prompt, a few well-chosen tools, output guardrails — often covers most needs.
Le risque du « Frankenprompt »
Anthropic describes an internal proliferation effect akin to Conway’s law: the architecture ends up reflecting the organizational chart. Each team adds its sub-agent, rules, and exceptions, eventually producing a "Frankenprompt," that is, a system prompt bloated with sometimes contradictory instructions. At this stage, complexity no longer improves the product; it simply makes it harder to maintain.
In an example cited by the company, Claude Code saw its system prompt reduced by over 80% on recent models, partly because skills and tools can now be loaded only when needed. Anthropic adds that models increasingly comment on their own progress and self-verify parts of their work, which can make some control layers less necessary than before.
Des sous-agents utiles, mais seulement dans certains cas
Anthropic does not say all sub-agents should be eliminated. The company finds this approach relevant when a single agent’s context window clearly degrades, when multiple tasks must run in parallel with separate contexts, or when the main agent struggles to choose the right tools. However, it advises against blind migrations. In a test on a support agent, consolidation did reduce latency but also broke a working process because the main agent started responding alone instead of consulting the right specialist.
To assist teams, Anthropic published a migration kit on GitHub offering a step-by-step method: identify existing patterns, test alternatives, then verify the most robust approach. The company also recommends limiting the number of tools exposed to each agent. According to it, tool selection quality tends to degrade beyond 30 to 50 definitions, which argues for skill libraries maintained by identified managers.
Les chiffres qui soutiennent le virage
The debate is not only theoretical. A Vals study cited by Anthropic concludes that agent teams can be faster than a single agent but without quality gains. Increasing from 10 to 30 or 100 agents brought no measurable improvement while significantly raising costs. Meanwhile, Eric Provencher, a developer at OpenAI, has already called agent swarms "almost certainly" wasted money.
For companies, the signal is clear: agentic AI does not automatically benefit from fragmentation. The most elegant architectures are not always those with the most sub-agents but those that maintain the right level of context, at the right time, with the fewest intermediaries possible.
Key points
- Anthropic recommends consolidating sub-agents when possible.
- The single-agent architecture ran about twice as fast.
- Claude Code reduced its system prompt by over 80%.
- Tool selection degrades beyond 30 to 50 definitions.
- Vals found no quality gain moving from 10 to 100 agents.
- Anthropic published a migration kit on GitHub.
By the numbers
- About 2x faster — Anthropic comparison between single-agent and multi-agent.
- Over 80% — reduction of Claude Code’s system prompt, according to Anthropic.
- 30 to 50 tools — threshold where selection degrades, according to Anthropic.
- 10, 30, and 100 agents — levels tested by Vals without quality gain.