Anthropic Tells Builders to Fold Subagents Into One Agent Loop

Anthropic's Applied AI team has reversed an assumption many agent builders have been shipping on for two years: more specialised agents do not automatically produce better results. In a guide published this week, the company says a single agent that loads context on demand through skills and tools frequently beat both monolithic prompts and orchestrator-plus-subagent designs across several enterprise deployments, often at lower cost and latency per task.

What the comparison actually showed

The most concrete number: a customer that ran its multi-agent harness against a single agent loop saw roughly twice the runtime speed from the simpler design. Another company had built a dozen specialist subagents inside one customer-facing agent. The result was a system that could not hold a consistent brand voice and added enough latency to hurt the experience.

The engineering reasoning is that every handoff between agents is a lossy translation. When a feature agent, a test agent and a review agent each get a partial summary of the work, the details that mattered get dropped. A single loop keeps the earlier steps in one context window, so the model writing the feature also writes the tests and can connect the results itself.

The org chart became the architecture

Anthropic describes the dominant failure mode bluntly: each team ships its own subagent, a platform team bolts an orchestrator on top, and the architecture quietly becomes a copy of the org chart. Every team then adds a sentence to the system prompt until the rules contradict each other, a condition the company calls the Frankenprompt. A clean prompt, four or five thin data tools and firm output guardrails, it says, cover most of what those sprawling graphs were built to do.

As evidence the harness was compensating for older models, Anthropic cut Claude Code's system prompt by more than 80 percent for advanced models, including the parts that existed to progressively disclose skills and tools.

When subagents still earn their place

Anthropic is explicit that this is not a blanket argument against multi-agent systems. Subagents remain the right call when one loop shows measurable degradation from context bloat, when tasks genuinely run in parallel with context windows that can be discarded afterwards, or when a single agent cannot pick the right tool. Folding is also not free: in one support-agent test the merged version cut latency but broke a working process, because the primary agent started answering on its own instead of consulting the specialist.

What builders should do next

The company shipped a migration kit on GitHub, anthropics/single-agent-loop-kit, that audits subagents one at a time and returns a verdict for each: fold, merge, call as a tool, or keep. It includes a demo on a mock system, since the audit is only as good as the evals behind it. Independent work points the same way: a Vals study found teams of 10, 30 or 100 agents delivered no quality gain over smaller setups while costs climbed.