Microsoft Research: Injecting NL Skills Cuts Agent Reasoning Cost by 6x
dair_ai · x · 2026-08-12
A new paper from Microsoft highlights that while reasoning modes excel at multi-step agentic tasks, they charge a 3 to 6x premium in output tokens compared to non-reasoning modes. Much of this spend is wasted on re-deriving procedures the model already figured out previously.
To amortize this cost, the researchers propose a method where:
- Skill Extraction & Injection: A coding agent reads a small corpus of existing trajectories, compiles them into a compact natural-language skill, and injects it into a non-reasoning model's system prompt.
- Performance: Across benchmarks like ALFWorld and tau-squared-bench, these skills recover 55% to over 100% of the reasoning gap.
- Efficiency: On two benchmarks, the skill-equipped non-reasoning model beats the reasoning mode outright, emitting 2.7 to 6x fewer output tokens and zero reasoning tokens.
More from coding & agent
- AI Coding Assistant Tuning Fail: Turning Codex into a Paper Machine — burny_tech · 2026-08-12
- Coinbase Payments Update Enables AI Agents to Transact Directly — kleffew94 · 2026-08-12
- Opinion: The Role of Compilers Will Change Heavily in the AI Era — pranjalssh · 2026-08-12
- gpt-5.6-sol Excels at iOS Development with Dedicated Mac and Simulator — dkundel · 2026-08-12
- Dev accidentally replaces OpenAI Codex with self-built tool Junior — zeeg · 2026-08-12
- LangChain: Only 7% of agent calls need frontier models, routing cuts costs by 74% — colinmcnamara · 2026-08-12