← sources

How do you decide what your agent is actually allowed to do? I'm building a permission check and want to know where my model breaks

source: reddit (reddit:AI_Agents)published 2026-11-18 14:52ingested 2026-11-19 02:35

retained text

Disclosure: I'm the founder of ORIMUSU and I built the tool I mention below. No links here, per the rules.

Most agent setups I've seen mix two things: what the model decides to do, and what it is permitted to do. I've been trying to separate them. Every operation the agent proposes gets checked against explicit permissions and evidence, and gets one of three outcomes: ADMITTED, DENIED, or UNRESOLVED. UNRESOLVED means "not enough evidence to decide," and it is not treated as a quiet allow.

It's an early Technical Preview. Right now it's a local Node.js CLI. The default mode only observes, and the blocking mode runs against a local mock broker. Real credential isolation isn't demonstrated yet.

Questions for people running agents for real:

1. How do you handle permissions today? Allowlists, human approval, sandboxing, or just trust?  
2. Is a three-outcome model useful, or would UNRESOLVED just get in your way?  
3. What would a check like this have to prove before you'd let it block your agent?

The project is open source (Apache-2.0). I'll share it in the weekly Project Display thread if anyone wants to look.

document fetched 2026-11-19 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xzh4xk/how_do_you_decide_what_your_agent_is_actually