# At what point should an agent stop relying on the model and let the harness take over? — retained record

> Retained reddit record from reddit:AI_Agents.

Canonical: https://highsignal.app/data/records/0f2f07cf2bec7c4a

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:AI_Agents)  published 2026-10-03 10:52 ingested 2026-10-04 02:35

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

## retained text



```
I've been building an agent runtime in Rust and I've started questioning how much of the "agent" should actually live inside the model.

The basic loop is easy:

model → tool call → result → model → repeat

Then the ugly cases show up.

The model repeats a tool call.
A command fails halfway through.
Context gets too large.
A previous operation needs to be rolled back.
The session needs to survive a restart.
The model says it's finished while the actual task clearly isn't.

My current approach is to keep reasoning in the model and move reliability/state management into the harness.

Things like retries, context compaction, loop detection, worktree handling, approval gates and recovery.

I'm still not sure where the boundary should be.

For people who have built agent runtimes or coding agents, what did you initially put inside the model that you eventually moved into the harness?

And what do you think should always stay model-driven?
```



document fetched 2026-10-04 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1wwkbmn/at_what_point_should_an_agent_stop_relying_on_the
