Across 27,280 Coding Agent Runs, 64% of Inference Spend Went to Failed Runs

Edward40 · hn · 2026-11-17

Runta's FrontierHarness eval v1.1 analyzes 27,280 coding agent runs and finds 64% of inference spend went to runs that ultimately failed. The post breaks down agent cost structure and discusses how early failure detection and better task routing could cut this waste — practical data for teams building agents and evals.

Original post →

More from coding & agent

coding & agent channel →