Putting an AI agent on real workflows: the failures that scared me
Front_Entrance259 · reddit · 2026-10-15
A practitioner building AI agents for real business tasks shares the scariest failure mode in production: each step looks fine in isolation, but the business should never have allowed the whole.
Case: A customer disputed a debt via text message; the next day the company's voice agent called them and ran the normal collection script, unaware of the dispute. Every step was individually correct; the whole was wrong.
The human-in-the-loop trap: The agent does 100 things in the time a person carefully checks 5. Approval gates either slow everything down or turn humans into rubber stamps — neither is real control.
The author asks how production teams handle this: hard approval gates on everything, checks outside the agent, or just hoping nothing breaks.
More from coding & agent
- Neon ships official AI Gateway adapter for TanStack AI with single-credential multi-model access — matei_zaharia · 2026-10-23
- Reflection AI's Beam agent auto-builds a Gemma-4 fine-tuning notebook, boosting accuracy 66.5% — danielhanchen · 2026-10-23
- Cua Spaces SDK lets Hermes and other agents run in sandboxes on your own hardware — Teknium · 2026-10-23
- PHP hit by two unauthenticated RCE flaws; ProjectDiscovery launches free internet-wide alert service — cyb3rops · 2026-10-23
- Modal open-sources Spindle, a Tinker-compatible training server that charges GPU time instead of tokens — josh_wills · 2026-10-23
- Testing the Azure Cosmos DB Agent Kit: what it found reviewing a Claude-built app, in RUs saved — adnan_hashmi · 2026-10-23