Human-in-the-loop isn't human authority: scoped grants beat click-approval fatigue
arthaudm · reddit · 2026-09-13
"Require human approval" sounds safe until every action needs a click: people stop reading and approval becomes muscle memory—the human is technically in the loop and practically absent. The author proposes scoped authority instead: each grant specifies who granted it, the exact action and target, limits and expiry, what evidence the agent must use, and what must return to a human. Routine actions run inside the boundary; anything outside stops. The tool enforces the grant rather than trusting the prompt to remember policy. The real design question isn't whether a human touches every action, but whether the agent can prove which human had authority for it.
More from coding & agent
- OpenAI Codex TUI's twinkling dot decoration gets copied into your text — doodlestein · 2026-09-13
- Clipboard-Automator: open-source ComfyUI nodes turn your OS clipboard into a zero-click pipeline — kastelan77 · 2026-09-13
- Dev ships obsideo-mcp, an MCP encrypted storage server agents can verify with signed possession proofs — Electrical_Offer5667 · 2026-09-13
- Prompt Injection Detector: a 6-point checklist for scanning untrusted text before it hits your LLM — blaizedsouza · 2026-09-13
- Trading agent's reflection loop was learning from noise: judging theses apart from PnL fixed half of it — ExplorerEconomy8233 · 2026-09-13
- A trick for efficient multi-agent runs: dispatch work, end turn, get woken up on completion — pvncher · 2026-09-13