What hard stops do you put on your AI agents? A practical red-line list
Virtual_Arachnid7916 · reddit · 2026-11-22
A developer building personal agents for shopping, email, research, scheduling, and meeting prep shares a non-negotiable hard-stop list before granting more autonomy: no purchases/payments without explicit approval; email drafts only, never sent as the user; no calendar accept/decline without confirmation; meeting notes reviewed before sharing; no permanent deletions; never typing credentials; and ignoring instructions embedded in web pages, emails, or documents to blunt prompt injection.
They ask the community what's missing, how to enforce limits (system prompts, tool permissions, approval steps, sandboxing), real cases where a hard stop saved or its absence hurt, and which stops were dropped for friction — with particular interest in defenses that hold against prompt injection rather than relying on the model "behaving".
More from coding & agent
- Sharing the Composer cache in GitHub Actions made PHPUnit integration tests 31% faster — DanielLockyer · 2026-11-22
- Google's Mirror Dimension: synthetic web for safe browser-agent testing — shashib · 2026-11-22
- Exposing Google/YouTube trending searches to agents via an MCP tool — Substantial-Tour-581 · 2026-11-22
- Stardock CEO shows Clairvoyance: hierarchies of 71 AI 'staff' agents building Sins of a Solar Empire II — draginol · 2026-11-22
- FieldCraft: pure-Rust open-source CAE tool bundles FEA, heat and fluids with MCP support for AI agents — Promptmethus · 2026-11-22
- Photo-to-game-asset workflow turns one photo into Unity-ready 3D plants for $4 — Promptmethus · 2026-11-22