Agents Still Burn 58% of Tokens After Failing
MaverikSh · reddit · 2026-07-14
A cited trace study reveals that in failed agent runs, about 58% of tokens are consumed after the model has already shown explicit signs of failure. This indicates the issue lies not just in model capability, but in the runtime control architecture.
The author argues that many cost controls are implemented only at the billing level, acting as post-event alerts that cannot stop a task from rapidly burning money. Truly effective approaches should:
- Reserve budgets at the run level rather than just checking single calls
- Use runid as a unified billing key so concurrent sub-agents share the same budget limit
- Set abort thresholds for repeated identical tool calls (e.g., stop if the same parameters are used 3 times consecutively) to avoid burning tokens in loops
The post specifically notes that concurrency worsens the problem: even if each sub-agent looks reasonable individually, the overall system can overspend without any single call crossing the line.
More from coding & agent
- Agent verified 'success' after double-voting: a lesson in harness vs model bugs — TejasKumar_ · 2026-10-26
- diginomica: agentic automation hype is overdone, and 'The Hackening' of enterprise security has begun — jonerp · 2026-10-26
- 'Vibe verification': ~60% of AI-written formal specs prove the wrong property — basedjensen · 2026-10-26
- Devin launches Security suite powered by Claude Mythos with $100M+ credits — DevinAI · 2026-10-26
- HackBench injects reward hacks into agent trajectories to stress-test monitors — DanielKhashabi · 2026-10-26
- Hugging Face ships hf CLI updates: deduped cache saves 19GB, faster buckets, retryable GPU jobs — yb2698 · 2026-10-26