Skip to content
Neural Daily

ToolsAnalysis

Anthropic: fold subagents into one agent loop with skills

🇭🇺 Olvasd el magyarul →

Anthropic's Applied AI team says a single agent loop with skills and progressively loaded tools often beats subagent designs in enterprise agents, on quality, cost and latency. It also released a migration kit.

By: Mikola Máté Patrik · · 3 min read

Anthropic illustration: a zigzag line connecting nodes inside a white lightning-bolt shape on a sage green background
Image: Anthropic · Press kit · Official article social image (og:image) · Source

The gist

On October 22, Anthropic published guidance on when to fold subagents back into the main agent loop (Anthropic). Comparing several enterprise deployments, its Applied AI team found that a single agent with skills often outperformed both the one-giant-prompt design and the subagent design on quality, often at lower cost and latency.

The usual failure mode, they say, is Conway's law: every team builds its own subagent, so the agent architecture mirrors the org chart. The result is lossy handoffs, conflicting prompt rules and evals optimized for subtasks. One customer saw roughly 2x faster runtime with a single loop than with its multi-agent harness.

The suggested design

  • One model, one context window, an open-ended tool loop, with limits (turn and cost budgets, write gates) set in the harness.
  • Each turn the model emits a batch of tool calls, the harness runs them in parallel and returns raw results to the same model.
  • The model writes the plan on the first turn and revises it as results arrive.
  • Domain procedure moves into skills, live data and actions into tools. Short descriptions are always visible, full content loads only when needed.
  • Above 30-50 tool definitions, tool choice degrades, so group tools by user intent rather than one per API.

When to keep a subagent

Keep one when a single loop measurably degrades from context bloat, when you need parallel work with disposable context, or when one agent struggles to pick tools or faces conflicting instructions. Even then, call it as a tool from the main agent. A full handoff only makes sense when the subagent must talk to the user directly.

How to migrate

Start with one shared end-to-end eval; even 10 golden examples will do. Then fold subagents in one at a time, starting with those that chain on each other's output or touch the same data. Don't just compare final scores: in Anthropic's test, one merge cut latency but broke a process that had been passing, and the break showed up in only one run out of two.

For the migration, Anthropic released the single-agent-loop-kit repo, which chains Claude API skills (build-eval, prompt-audit, hillclimb, cost-optimize) into a guided workflow with a demo system. If you are designing an enterprise agent now, start from one loop and split it only when measurements say so.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles