A commenter found a hole in our agent approval gate: 408 untracked files and 138 agent commits later
Vektoris_AI_Workflow · reddit · 2026-10-17
A postmortem from a two-human, multi-agent team: they claimed "agents write, but a human approves every change via git diff." A commenter asked whether other agents read proposed changes while pending. Checking revealed the gate was hollow.
- Approval was theater: approvals happened in chat, commits came late or never—50 modified and 408 untracked files in the working tree, 8 decision notes never committed, the oldest open for 5 weeks.
- Everyone reads from disk: the morning briefing, drafting agent, and finance agent all used whatever was in the folder, approved or not; the search index rebuilds after commits but walks the folder, so unapproved notes got indexed anyway.
- The reverse hole: the ingest agent had always been allowed to commit itself—138 of 150 commits since September. "Committed" never meant a human looked. Of 73 needs-review flags, none had an owner; oldest dated to June.
- Deliberately small fixes: no hard read gate (agents would use stale prices); label unreviewed state; briefings list critical open files (prices, decision notes, client pages, invoice register) and drafts citing them say so; weekly review handles items open >7 days; new flags require an owner; no-review commits must record base commit + file hash with checks on merged result.
The authors admit none of this prevents anything—it only works if someone looks—and ask how others handle review-branch drift and write- vs read-side gates.
More from coding & agent
- A one-shot cron job prompt got Claude to crack ciphers and advance a 1932 number theory bound — IgorCarron · 2026-10-17
- Backend migration from Python to Go now takes one engineer just one week — rakyll · 2026-10-17
- Audio founder says Claude one-shot decompiled rival algorithms, moat gone — ChrisGPT · 2026-10-17
- Google's playbook for agent evals: four layers of checks beyond the vibe check — rseroter · 2026-10-17
- AgentForesight: a 7B online auditor that catches multi-agent failures mid-run — jiqizhixin · 2026-10-17
- Qwen3.8-27B shows a weird write-delete loop in /tmp before every web research task — Training-Respect8066 · 2026-10-17