--- id: "2026-10-22-anthropic-fold-subagents-single-agent-loop-guide" url: "https://postcutoff.com/e/2026-10-22-anthropic-fold-subagents-single-agent-loop-guide/" as_of: "2026-11-22T13:32:00+01:00" date: "2026-10-22" date_precision: day category: agents importance: 2 confidence: high status: [Confirmed] sources: 2 editor: human editor human_review: null version: "2026-10-29" --- As of: 2026-11-22 13:32 CET. Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/ Canonical page: https://postcutoff.com/e/2026-10-22-anthropic-fold-subagents-single-agent-loop-guide/ # Anthropic advises folding subagents into one agent loop with skills and tool search, and releases a reference kit to measure it On Oct 22, 2026 Anthropic published "Building agents: When to fold subagents into your main agent loop" (Diane Sarkis), arguing that with newer models most router-plus-subagent systems should become one agent loop that loads skills and tools on demand, keeping a subagent only where evals show it pays. A companion repo, anthropics/single-agent-loop-kit (Apache-2.0), walks teams through measuring the change. ## Key facts - Article (claude.com, Oct 22, 2026, author Diane Sarkis): replace subagents where possible with one loop using progressive disclosure (skills for procedures, tool search for tools); keep a subagent only for measured context bloat, parallel work that benefits from fresh context, or tool-selection problems and conflicting instructions - Numbers cited: one customer's single loop ran about 2x faster than its multi-agent harness; Anthropic cut Claude Code's system prompt for more advanced models by over 80%; tool choice tends to degrade beyond 30-50 tool definitions in the prompt - Method: change one thing at a time against a task-completion eval, starting from about 10 golden examples, and run tasks more than once (in a sample support agent a failure appeared in one of two runs) - 'Each new model lets you move more of the decision-making out of the harness and into the model itself.' - Kit: github.com/anthropics/single-agent-loop-kit (created Oct 15, Apache-2.0, ~420 stars by Oct 29), a reference implementation 'not maintained and not accepting contributions'; steps measure, combine, fold, with a runnable support-desk example (a router and six subagents) ## What happened Anthropic's applied team published guidance that runs against the multi-agent fashion of 2025–26: as models improve, a single agent with on-demand skills often beats an orchestrator with specialist subagents. The companion kit scores a system before and after each subagent is removed. ## Why it matters It is first-party advice from a frontier lab on agent architecture, and it explains why harnesses are getting thinner as models take over decisions that used to be hard-coded. ## Your AI and this story - GPT-6.2 Astra (training cutoff May 2026): 144 days after its cutoff - Claude Opus 5.5 (training cutoff June 2026): 114 days after its cutoff - Gemini 4 Hydrogen (training cutoff March 2026): 205 days after its cutoff - Grok 4.7 (training cutoff May 2026): 144 days after its cutoff ## Sources 1. [Claude: Building agents — When to fold subagents into your main agent loop (Oct 22)](https://claude.com/resources/articles/building-agents-when-to-fold-subagents-into-your-main-agent-loop) (claude.com, official) 2. [GitHub: anthropics/single-agent-loop-kit](https://github.com/anthropics/single-agent-loop-kit) (github.com, code) ## Changes - 2026-10-29 (filed): Created (news agent; a week after publication) ## Related - 2026-10-29: [Spawn's founder says ten people and a fleet of Claude Fable agents ship ~108 changes an hour, spending 30x payroll on tokens](https://postcutoff.com/e/2026-10-29-spawn-fable-agent-fleet-30x-payroll-tokens/index.md)