Decision models for agent tool selection: 50 tools to 3, 47% fewer tokens, 22% faster
rajistics · x · 2026-11-13
Inspired by Michelle Chen's aiDotEngineer talk, the author tested decision models for tool selection in coding agents: picking 3 needed tools out of 50 GitHub tools cut token usage by 47% (229K → 121K) and latency by 22% (18.2s → 14.2s), with the selector itself taking only 0.48s for 10 questions in one call. OpenHands' John-Mason added SDK support for this. A concrete demonstration of how decision models can slim down coding-agent contexts.
More from coding & agent
- LangChain raises $300M at $3B valuation, with 7,000+ companies on LangSmith — LangChain · 2026-11-14
- Largest systematic agentic eval campaign launches: test agents as tool users, not tools — animesh_garg · 2026-11-14
- Perplexity agent orchestrated 48 parallel sandboxes to render photoreal Pantheon video in 3 hours — AravSrinivas · 2026-11-14
- Cat Wu, ex-Claude Code product lead now at OpenAI: 10 great evals beat hundreds — aakashgupta · 2026-11-13
- AsyncGRPOTrainer now stable in TRL v1.18.0: generation and training on separate GPUs — SergioPaniego · 2026-11-13
- YC-backed Reason claims frontier coding performance at up to 20x lower cost — Scobleizer · 2026-11-13