1,360 agent approval gates in 30 days, only 55 needed a human: a tiered oversight setup
ldrx · reddit · 2026-11-12
The author runs long-lived Claude Code sessions on a shared task board building a real open-source project. In 30 days: 1,360 approval gates — 678 auto-cleared, 475 handled by another agent, only 55 needed the human (17 involving credentials), 152 unanswered. Rule of thumb: agents decide routine matters; spend, keys, destructive actions and brand calls need human sign-off. Completed tasks go to an independent reviewer session that sees only the task and diff — it rejected 45% on first pass (failing tests 41%, unclear writing 22%, no proof it works 8%, and more). After two rejections, agents escalate to the human.
More from coding & agent
- Rails 8.2 adds read-only query command for AI agents; Rust-in-monolith shows Rails is thriving — jonmarkgo · 2026-11-12
- Indie dev uses his own BizPeel MCP to research marketing for BizPeel itself, surfacing 4 copyable growth plays — yihui_indie · 2026-11-12
- Blender MCP adds World Labs Marble integration to generate 3D worlds inside Blender — sidahuj · 2026-11-12
- ffmpeg-skill: Open-Source Local Video Editor for Coding Agents Hits 5.2k Stars — adnan_hashmi · 2026-11-12
- Anthropic's Fable 5.2 builds production-grade apps unattended — one dev shipped a video editor in 3 days — every · 2026-11-12
- AI agent finds sponsor leads via People Search MCP: 10 contacts for under $2 — cneuralnetwork · 2026-11-12