Microsoft's TokenEconomizer cuts coding agents' recurring token costs with a black-box controller
MicrosoftResearch · hf · 2026-10-20
Microsoft Research, UC Berkeley and Microsoft Security introduce TokenEconomizer, an experience-driven external controller that reduces recurring, environment-specific token costs of coding agents across sequences of real-world tasks.
- Motivation: agents' token efficiency is usually evaluated on isolated tasks, but real developers use them to continuously build and maintain large-scale projects, where recurring environment-dependent token costs emerge that existing methods struggle to reduce
- Method (black-box): learns (i) which operational micro-skills to inject, (ii) which persistently irrelevant tools to prune, and (iii) when to intervene in unproductive failure loops
- Authors include Ion Stoica, Michel Galley, Bin Yu and Jianfeng Gao; evaluated with Claude Code, Codex and GitHub Copilot CLI
- Code open-sourced on GitHub
More from coding & agent
- REA gains 214K stars in a week, tops GitHub's fastest-growing AI repos — sven_ai · 2026-10-20
- Dev goes 'graph pilled': graphs will anchor every company's world model — _AustinCalvert_ · 2026-10-20
- RuneBench: Benchmarking AI Coding Agents by Playing RuneScape — charliermarsh · 2026-10-20
- A 5-step loop for debugging AI agents in production: trace everything, watch for silent failures — Tracify · 2026-10-20
- New Claude Cookbook: Auto-Review Agent Session Transcripts for Security Issues — brada · 2026-10-20
- Conversational Video Editing: Driving CapCut Web with Codex for Marketing Videos — Div_pradeep · 2026-10-20