Microsoft's TokenEconomizer cuts coding agent costs 18–31% by learning from past sessions
Microsoft Research · rss · 2026-11-17
TokenEconomizer, an open-source plugin from UC Berkeley and Microsoft Research, learns from an agent's past sessions in a repository to cut future costs without modifying the agent or model.
- Waste profile (Claude Code on SWE-Bench): 30% of tool calls repeated earlier-task actions after 5 tasks; 16 of 23 loaded tools were never called; the 17% of failed tasks consumed 32% of spend
- Three mechanisms: micro-skills (one-line repo notes injected after multi-session confirmation), tool pruning (drop unlikely tools, keep core ones), failure-loop intervention (judge model detects stuck agents and suggests new approaches)
- Results: 18.4–30.9% average cost reduction across 277 SWE-Bench tasks and 5 agent-model combos, with no performance loss; 16.0–21.2% savings on Web-Bench; cumulative costs beat default by the sixth task
- Lowering reasoning effort saved similar amounts but degraded performance substantially
More from coding & agent
- llms.txt Protocol Goes Mainstream as Google Antigravity Adopts It Two Years On — jeremyphoward · 2026-11-22
- He built a feed only agents can read: no UI, semantic matching, agent-rated posts — imdogefather · 2026-11-22
- Claude Code Debuts Experimental Code Mode as AI-Driven Tooling Gains Momentum — lxfater · 2026-11-22
- ClaudeAI Weekly: Single-HTML-File Arms Race, 8.9B-Token Strawberry, Post-IPO Anthropic — ClaudeAI-mod-bot · 2026-11-22
- Agno's AgentOS turns any agent into a production FastAPI app with MCP, RBAC and tracing — pritisinghhhh · 2026-11-22
- Chinese devs vent: app store rejections and copycats are just the norm — tinyfool · 2026-11-22