Failed AI Agents Consume 40% More Tokens Than Successful Ones
0xsachi · x · 2026-08-18
Data from 13,000+ OfficeQA runs reveals an early warning sign for AI agent failure: incorrect agents take up to 50% more steps, consume 40% more compute, and incur 40% higher costs per episode compared to successful ones. Failed trajectories not only take longer but waste significantly more compute, tokens, and money. When an agent takes more actions without making real progress, it is a useful indicator that it is heading towards failure.
More from coding & agent
- /improve-threejs：把 vibe coding 的 Three.js 垃圾变流畅 — aidenybai · 2026-08-18
- Dev trick: Visualizing codebases into interactive diagrams with Claude — round · 2026-08-18
- Testing MCP Apps with three contracts: jsdom, Playwright, and manual checks — philrox_ · 2026-08-18
- New research: are LLM agents time-aware? Can they estimate task duration? — maksym_andr · 2026-08-18
- OpenAI's steipete on coding agents expanding into observability and long-running ops — steipete · 2026-08-18
- SQL MCP Server: connecting AI agents directly to SQL data — adnan_hashmi · 2026-08-18