What actually has the authority to stop your agent when it goes wrong?
No_Progress92 · reddit · 2026-08-31
The author argues agents are improving fast at planning, tool use and multi-step workflows, but control remains underdeveloped: most setups rely on prompt instructions, soft harness constraints, or after-the-fact review, which become unreliable once agents have real access to tools, files and systems. He is building an open-source approach that treats control as a separate runtime boundary — identity, permissions, validation, and an audit trail the agent cannot rewrite — and asks how others handle having real authority to stop a misbehaving agent.
More from coding & agent
- Quote: Agents work worse the further they are from human supervision — danshipper · 2026-08-31
- LLM coding trap: 95% efficiency leads to 5% total paralysis — burny_tech · 2026-08-31
- Why self-improving agents never ship: a memory architecture with mandatory rollback — inbask · 2026-08-31
- Developers debate: Would you trust AI coding agents in YOLO mode? — NoMoreHappyPath · 2026-08-31
- OpenClaw 2.0 lands with 16,000+ PRs, multiplayer sessions and a rebuilt browser app — The Decoder · 2026-08-31
- Planning to fork Octo.nvim to build a customized plugin — 4310sy · 2026-08-31