An edit apply model is a small model used in a two-model code-editing setup: a large model decides what to change and describes the edit as a Markdown explanation, and the small “edit apply” model then rewrites the entire file to carry that explanation out. As described in Don’t Build Multi-Agents, the idea rested on the observation that, at the time, getting a small model to rewrite a whole file from a description was more reliable than getting a large model to output a properly formatted diff.
Usage
That post describes the pattern as a common practice in 2024 among coding agents, IDEs and app builders, including Devin, at a time when many models were bad at editing code. It uses the pattern as an example of what goes wrong when decisions and actions are split across models: the small model would often misinterpret the large model’s instructions and make an incorrect edit because of the slightest ambiguity in them. The post adds that edit decision-making and applying are now more often done by a single model in one action.
Related Terms
- Context engineering — the post discusses the pattern among its real-world examples of architecting agents to avoid conflicting decision-making