Anthropic: One AI agent now beats fleets of subagents

Anthropic reports that a single AI agent with on-demand context often beats multi-agent setups on quality, cost, and latency — and in one comparison, the single-agent architecture ran about twice as fast as the multi-agent version. The company published a migration kit on GitHub to help enterprise teams consolidate their subagents rather than keep building them.

The stakes are practical for any company that has already deployed agents in production. Multi-agent designs promised division of labor; Anthropic's Applied AI team now says they often deliver higher latency, rising costs, and information loss at every handoff.

Why does one agent now beat many?

The answer is model capability. Newer models — starting with Opus 4.5, along with Opus 5.1 and Fable 5.1 — can manage complex plans directly in a single loop, according to Anthropic. That lets one primary agent load context only when needed instead of routing work across a fleet of specialized subagents.

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 ran a dozen specialized subagents inside its customer-facing agent; the result, Anthropic says, was an inconsistent brand experience and poor usability from high latency.

Much of the scaffolding around older agent designs has become redundant. Anthropic describes a familiar pattern: intent detection, a fixed processing pipeline, user segmentation, separate recommendation branches, and an orchestrator sitting on top. The model can now handle most of that work itself for a given use case.

What is 'Conway's Law' for agents?

Anthropic calls the failure mode "Conway's Law" for agents: the architecture becomes a copy of the org chart. Each team ships its own subagent until hundreds of overlapping agents pile up with no shared standards. What follows is what Anthropic calls a "Frankenprompt" — every team adds rules to the system prompt until the rules contradict each other.

The fix, according to the company, is restraint. A clean system prompt, four or five lean data tools, and fixed output guardrails cover much of a typical use case.

Anthropic applied this to its own products. 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 makes some extra control layers unnecessary.

When do subagents still make sense?

Anthropic does not recommend merging everything. Subagents remain justified when:

• a single agent's context window shows measurable performance degradation;
• multiple tasks need to run in parallel with separate context windows;
• a single agent struggles with tool selection.

The company also warns against merging blindly. In one test with a support agent, folding in a subagent reduced latency but broke a previously working process: the primary agent started answering on its own instead of consulting the right specialist.

The GitHub migration kit walks teams through a workflow to identify existing patterns, evaluate alternatives, and verify the best approach before committing.

How should teams manage skills and tools?

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.

There is a hard operational limit: the model's tool selection tends to degrade beyond 30 to 50 tool definitions. Designated owners on each team should maintain the skill libraries to keep them from ballooning.

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

Is Anthropic alone in this finding?

No. 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.

That alignment matters beyond engineering aesthetics. Companies have spent two years building orchestration layers and agent fleets around models that could not hold a complex plan in one context window. If Opus-class models can now do the orchestration themselves, the economics of that stack change: fewer handoffs means less information loss, lower token spend, and the roughly two-fold latency improvement Anthropic measured in its comparison. Teams that audit their subagent sprawl now — with the migration kit as the checklist — will likely ship faster and cheaper than those that keep rebuilding the org chart in software.