The Missing Harness: Why Agent Loops Need More Than a Retry Prompt
Building Autonomous Goal Loops That Deliver
Autonomous agent loops that simply retry until tests pass fail when the goal is to grow an unknown product capability. This article argues that agents need a harness that exposes real failures, locates missing capabilities, and preserves lessons across sessions. It details a three-loop system with a deterministic floor, a direction-finding scorer, and an authority model that prevents self-grading. The key insight: the repository, not the prompt, is the control plane, and people must own the rounds while agents own the turns.
A good round leaves a working capability, evidence that it works, and a harness that will not pay for the same lesson twice.
- madamelic
I do sometimes wonder if we are all writing more than we would if we just wrote the code ourselves.
I am asking that partially as a rhetorical question but also wondering your thoughts on when to deploy 'loop engineering' versus 'one-shot' versus writing the code yourself.
Additionally the post seems to talk in broad strokes without a specific proposal on a 'scorer' / determining adherence to the state goal. Do you have a thought on how that should be expressed? Do you feel that a human in the loop slows things down unnecessarily?
- Oscalemor
Any data pre/post implementation feels like everyone wants to optimise this. Caught myself in this too; how do we actually measure it
- TooTony
Sorry, the animated banner at the top makes it hard for me to focus on reading.