UCLA paper: ~18% of AI agent output tokens are removable waste, pruning 41-46% of reasoning barely hurts
rohanpaul_ai · x · 2026-11-09
A UCLA/Northeastern paper, "How Do AI Agents Waste Your Money?", introduces a unified framework for measuring token waste across heterogeneous agent systems.
Method: record and normalize full agent trajectories, use LLM-as-a-Judge to mark removable spans, map them back for token-level measurement, then validate by removing spans, resuming execution, and measuring task outcomes.
Key findings:
- About 18% of agent output tokens are removable waste (repeated policies, refetched files, dead ends); cutting them barely hurts success rates
- Online pruning removes 41–46% of model reasoning while largely preserving performance across two agent benchmarks
- Waste varies more across tasks and agent settings than across models; different harnesses introduce distinct inefficiencies even with the same model and task
The framework turns token waste into an actionable optimization signal — directly relevant to cost reduction for token-billed agent apps.
Related event: Study: About 18% of AI Agent Output Tokens Are Waste(2 posts)→
More from coding & agent
- Garry Tan: Apple should rebuild its software teams with hundreds of codegen agents — illscience · 2026-11-09
- Why top coding models without dedicated s-tier harnesses go untested: the harness-model pairing effect — illscience · 2026-11-09
- WonderDuck: open-source robot duck ships full sim-to-real pipeline — philfung · 2026-11-09
- AI Engineer Code Summit kicks off in SF with agent-heavy agenda — threepointone · 2026-11-09
- Agent Security Incidents Need Tamper-Proof Audit Trails, Khala Research Report Argues — 0xSammy · 2026-11-09
- SenseTime open-senses Code2Video: an open model that makes editable videos via code — HeyZoyaKhan · 2026-11-09