AI Agents Waste ~18% of Output Tokens, Paper Finds; Pruning 41-46% Barely Hurts Success
rohanpaul_ai · x · 2026-11-09
A paper from UCLA, UIUC, Northeastern and UT Austin introduces a unified framework for measuring token waste in AI agents: a proxy records full trajectories, an LLM-as-a-Judge marks removable spans, and deleting them while resuming execution verifies nothing breaks.
- Across 6 frontier models and 4 benchmarks, 18% of output tokens were waste (repeated policies, refetched files, dead ends).
- Online pruning removed 41-46% of reasoning while largely preserving task performance; pruning Kimi-K3's reasoning on SWE-bench Verified cut 41% with success at 89.5% vs 89.0%.
- Waste varies more across tasks and harnesses than across models; different harnesses introduce distinct inefficiencies under the same model and task.
Takeaway: to cut agent costs, find where your task and harness create waste instead of just capping tokens.
Related event: Study: About 18% of AI Agent Output Tokens Are Waste(2 posts)→
More from coding & agent
- Open harness eval: Hermes+Opus 5.5 tops office-task benchmark at 78.8% pass rate — NousResearch · 2026-11-09
- DeepSeek open-sources Super-Agent, a coding agent that watches Feishu chats and ships PRs — dl_weekly · 2026-11-09
- Context engineering field notes: poisoning, distraction, confusion and clash — sharpeye_wnl · 2026-11-09
- Harness 2025 vs today vs 2027: from function calls to swarms and async threads — blaizedsouza · 2026-11-09
- Claude Opus turns photos into 3D Blender scenes via Python, judges itself by pixel match — felixrieseberg · 2026-11-09
- SnapWidget ships an MCP server: Claude or ChatGPT embeds Google reviews on your site — hermanschutte · 2026-11-09