What Actually Broke When We Put Agents on Real Back-Office Work
TapNorth0888 · reddit · 2026-11-02
The author shares lessons from running agents on real back-office ops (invoices, chasing missing docs) where tutorials fall short:
- Unknown outcomes: a timeout is not a failure. If the agent may already have written to the ERP or sent an email, treating "no response" as "retry" creates duplicates. Every outward action is now forced into done / not done / unknown, and unknowns go to a human.
- Review queues without context: dumping failures into a human queue left reviewers blind. Each handoff now carries the fields, the abstention reason, and what to check — and those human corrections become the eval set.
- Credentials: one shared API token for "the bot" made it impossible to revoke a single agent. Each agent now gets its own scoped token, named owner, and expiry.
The author asks what broke first for others once agents left the happy path: state, ownership of side effects, or something else.
More from coding & agent
- Dev's hands-on with Kiro Crew: spec-driven agent workflows yielding hundreds of automated tests — DigitalColmer · 2026-11-02
- ripgrep searches 30k Obsidian notes in 0.15s vs 4s with built-in search — dSebastien · 2026-11-02
- TAWA GUARD Launches Pre-Model Input Checks to Block Hidden Prompt Injections in Agent Inputs — TawaGuard · 2026-11-02
- Beehiiv price hike alternative: email platforms ranked for agent/MCP workflows — ayushtweetshere · 2026-11-02
- AiVoiceBus open-sourced: one Go binary message bus for AI coding agents across machines — baditaflorin · 2026-11-02
- Researchers Swap Lab Notebook Tips for Managing PhD Projects and AI Agents — 3scorciav · 2026-11-02