They Added a No-Progress Kill for Agents — Cost Fell 19%, But Cost per Successful Task Rose
Specialist_Dish_9087 · reddit · 2026-09-12
- A team added a no-progress kill (3 identical tool calls → stop). Total bill dropped 19%, but cost per successful task went up, two months running.
- Why: runs that used to flail for 30 minutes then succeed carried all that flailing in their cost; now they stop at 4 minutes, so money still goes out with nothing to divide it by.
- Someone suggested turning the kill off to improve the metric — a red flag that the KPI incentivizes bad behavior.
- They now report two numbers: unit cost, plus share of monthly spend on zero-output runs (a quarter last month).
- Open problem: user-aborted runs, verifier rejections, and runs hitting the ceiling while still progressing all look identical in tagging — deciding which flailing runs matter remains unsolved.
More from coding & agent
- uv rewrites ZIP extraction inspired by Bun, Jupyter installs now 25-30% faster — mohamedmansour · 2026-09-12
- Developer's Codex Agent Just Made Its First Real Purchase on Its Own — Rasmic · 2026-09-12
- Using the Open Knowledge Format to Give Agents Long-Term Context Across Sessions — rseroter · 2026-09-12
- GPT 6 Astra rewards strong theory of mind, early user says — brandon_galang · 2026-09-12
- Bezalel Packs Memory, Email, Payments and More Into One MCP URL for Any Agent — Rasmic · 2026-09-12
- Neon ships Claimable: agents provision Postgres without signup, humans claim later — matei_zaharia · 2026-09-12