Claude Duet's stop rules show approval fatigue is the hard part of proactive agents
shoumikgoswami · reddit · 2026-10-31
Anthropic's research preview of Claude Duet ships with published stop rules that address the question most agent launches skip: when does an autonomous agent stop and ask?
- It reads, drafts and prepares work autonomously, but must show and await approval before sending messages, paying, signing, or deleting anything.
- Some actions are forbidden even with approval: trading securities, moving money between accounts, accessing someone else's account.
- Payments, order deletion and account creation are exempt from "don't ask again" rules no matter how often approved; scheduled background runs face the same gates.
- Anthropic calls approval fatigue "a serious issue" — too many requests train users to tap through — and will keep tuning interrupt frequency.
- A separate automated check runs outside the working model to resist prompt injection hidden in pages, but the write-up admits it "will occasionally miss something or flag a harmless message".
The poster's suggested takeaway: judge a proactive agent by its stop rules, not its feature list. Note this is Anthropic's own description, not an independent test.
More from coding & agent
- Microsoft's ScholarEvolve evolves agent harnesses with frozen models, +14 points on AppWorld — jiqizhixin · 2026-10-31
- Team uses Gemini as a subagent for tricky architecture and system design problems — AccBalanced · 2026-10-31
- Toska: Open-Source Krita Fork With an Agentic AI Painter for Under $7/Month — yawk-ra · 2026-10-31
- 'Don't use tools you can't fully control': OpenClaw pitch sparks agent autonomy debate — heyneighbor · 2026-10-31
- Dev builds Waifu Weather iOS app entirely with Claude Code, from Blender 3D modeling to music — MurkyGiraffe5879 · 2026-10-31
- Open-source /REA-growth skill reverse-engineers any company's growth playbook for $0.58 — _AustinCalvert_ · 2026-10-31