Anthropic's new guide: fewer subagents, single-loop agents 2x faster, Claude Code prompt cut 80%
emarkosov · reddit · 2026-10-24
Anthropic published an 18-minute guide (Oct 22) arguing most agents should ditch extra subagents in favor of one model, one context window, and one open-ended tool loop:
- Domain procedures go into skills; hard limits and lookups into tools; let the model write and revise its own plan.
- Keep subagents only when context bloat measurably degrades performance, you need parallel disposable-context work, or too many tools/instructions cause confusion.
- One customer's single loop ran 2x faster than its multi-agent harness; several deployments saw better quality at lower cost and latency.
- Tool choice degrades past 30-50 tool definitions in the system prompt — group by user intent and use tool search with deferred loading.
- Anthropic cut Claude Code's own system prompt by over 80% for newer models.
- Build evals before refactoring: 10 golden examples suffice, and run each task more than once to catch regressions.
- Migration kit on GitHub (anthropics/single-agent-loop-kit) with skills like build-eval, prompt-audit, hillclimb, cost-optimize.
The poster confirms from experience: six subagents shrank to two — most were just skills with a separate context window.
Related event: Anthropic Advises Folding Subagents into Main Loop for Faster Agents(3 posts)→
More from coding & agent
- Claude Code audit finds 214,856 "permanently deleted" photos lurking in Android hidden cache — Loo_Atreides · 2026-10-26
- Someone sent the first customer fax via ChatGPT over Onetime Fax's new MCP server — tdhopper · 2026-10-26
- China Telecom's XingChen lab open-sources Xing4.0-29B-Intent, a 29B MoE with 4B activated params — FellMentKE · 2026-10-26
- Codebases sized to fit one context window dramatically improve agent outcomes — julianharris · 2026-10-26
- 42-Minute Video Breaks Down Why AI Agents Fail: Mostly Not the Model's Fault — JohnSavill · 2026-10-26
- Giraffe: Open-Source Baseline Testing for OpenAI-Compatible LLM Endpoints — ImmediateBig8479 · 2026-10-26