# Where should the approval boundary sit when an AI agent can extend its own capabilities? — retained record

> Retained reddit record from reddit:AI_Agents.

Canonical: https://highsignal.app/data/records/5d052686c7d9501d

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:AI_Agents)  published 2026-11-11 17:01 ingested 2026-11-12 02:37

[original source ↗](https://www.reddit.com/r/AI_Agents/comments/1xtn4mu/where_should_the_approval_boundary_sit_when_an_ai/)

## retained text



```
Where should the approval boundary sit when an AI agent can extend its own capabilities?

I've been working on the architecture of a repository-based AI agent, and one question has become particularly important: how much freedom should an agent have to improve its own capabilities?

I don't think an agent should be able to silently modify its working environment, install new tools, and decide for itself that everything is safe.

The approach I'm developing is a little different:

The agent identifies a missing capability, develops a candidate tool in an isolated sandbox, tests it there, and saves the exact tested version for human review. Only after approval can that specific version become part of the working system.

The agent can propose improvements and do the technical work, but it cannot approve its own changes or move them into production.

What I'm still thinking about is the approval boundary itself.

Is reviewing the code and test results enough? Should approval also be tied to a specific version hash, permissions, dependencies, and runtime restrictions?

And what happens when an apparently harmless modification changes the behavior of something that was already approved?

I'm interested in how others approach this problem. There seems to be a difference between giving an agent the ability to improve and giving it unrestricted authority over its own environment.
```



document fetched 2026-11-12 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xtn4mu/where_should_the_approval_boundary_sit_when_an_ai
