I audited 155 agent jobs: the costliest failures all reported success
ranbuman · reddit · 2026-08-18
The author reviewed 155 delegated jobs across projects; 14 failed, and not one due to the model misreading the task — 11 were 400-900s timeouts, plus one DNS failure, one 529, and a session limit on the far side.
What actually cost days was the class of tools that return success while doing nothing: a browser fill reporting ok with an empty field, getstate always reporting an empty textarea, and a form whose submit click did nothing because g-recaptcha-response was empty. The agent then continues with a false belief for the rest of the run.
Fixes:
- Assert on the effect, not the return code
- Read the effect back through a different path than the one that made the change (e.g. DOM property vs. rendered accessibility node — the bug usually lives in only one)
- Carry the dead session's id in the error so work can resume instead of restart
The failure log he expected to be full of bad reasoning was full of infrastructure.
More from coding & agent
- The right version of AI creative workflow automates tedious work, not taste — umesh_ai · 2026-08-18
- Pilot Harness: An Out-of-the-Box Desktop Client for DeepSeek Harness — op7418 · 2026-08-18
- Neo4j launches Agent Memory: A graph-native memory system for AI agents — techNmak · 2026-08-18
- Grok 4.6 Review: Faster, better value, strong at frontend coding — prasenx · 2026-08-18
- Best practices: Making your landing page agent-ready with Markdown and llms.txt — handotdev · 2026-08-18
- Live Experiment: Can an AI Own a Knowledge Library? — dfinke · 2026-08-18