---
title: "Corrective feedback should be git-diffable"
url: https://recursiveai.org/research/git-diffable-feedback/
description: "When an agent misbehaves, the durable fix is a short, readable rule appended to its definition. The learned policy is text — you can read it, diff it, blame it, and revert it."
date: 2026-08-27
author: Kurtis Welch
---

# Corrective feedback should be git-diffable

August 27, 2026 · Kurtis Welch

When an agent misbehaves, there are three places the fix can live: in the model (retrain or fine-tune), in the moment (correct it in conversation and hope), or in the definition. We settled on the third, and the mechanism is almost embarrassingly simple: **a short, readable rule appended to the agent’s definition file, written by a refiner agent that reads the live transcript.**

The properties this buys are the properties text has always had:

- **You can read the policy.** The agent’s accumulated corrections are a list of plain-language rules, not weights. Anyone can audit what the agent has been taught.
- **You can diff it.** Every behavioral change is a commit. “Why did the agent start doing X?” has a `git blame` answer.
- **You can revert it.** When a correction stops being true — the upstream bug got fixed, the style guide changed — you delete a paragraph instead of retraining a model.

The refiner reads the transcript where the misbehavior happened, drafts the narrowest rule that would have prevented it, and appends it. Rules accumulate; stale ones get pruned in review. The refiner has been refined by itself, seven entries deep — the mechanism is stable under self-application.

The second-order effect matters more than the mechanism: **this is what makes agents teachable by non-programmers.** One of our users runs a paid consulting practice on agents she corrects entirely in conversation — the agent’s own prompt tells it where its knowledge files live and instructs it to edit them on request. A correction given once lands in a file and applies to every output thereafter. The high-stakes phrasing rules her domain requires — the kind a general assistant gets wrong every time and cannot durably learn — were taught once, in plain English, and have held since.

Fine-tuning has its place. But for the long tail of “don’t do that again” — which is most of what teaching an agent actually is — a diffable text rule beats a gradient update on every axis that matters operationally: latency of the fix, auditability, reversibility, and who is allowed to do the teaching.