Will the tricks an AI agent learns while coding also make it better at tasks that have nothing to do with writing code?

This explores whether improvements agents pick up on coding tasks carry over to non-coding work, or whether they only help with more code. The short answer from the corpus: there is strong evidence that these improvements travel across models and programming languages, some evidence that they reach neighboring technical fields, and little direct evidence that they reach clearly non-coding domains. The most-cited transfer result is Sakana AI's Darwin Gödel Machine. Improvements it found to its own tools and workflows while training on Python carried over to Rust, C++, and Go, and to different foundation models Do agent improvements discovered in one model transfer to others?. That is real portability, but every one of those destinations is still a coding task.

The furthest reach in the corpus is AIDE2. Its gains held up on four benchmarks it had never been tuned on, including physics-based weather forecasting, which lay outside the tasks used to select its improvements Do AIDE2's improvements transfer to unseen tasks?. Weather forecasting is a scientific domain, but the agent still does the work by writing and running code. Harness work shows a similar pattern. A fixed runbook built around frozen models lifted scores on Terminal-Bench across several models without any retuning Can execution harnesses lift model performance without retuning weights?. That shows the improvement isn't tied to one model. It doesn't show it works outside the terminal.

The corpus suggests a different way to see the question. Code may not be a domain at all. It may be the medium an agent thinks through. One note argues that code is uniquely executable, inspectable, and stateful, which lets it hold an agent's reasoning, its actions, and its checks in one loop Can code serve as the operational substrate for agent reasoning?. If that's right, a coding-born improvement transfers whenever a non-coding task can be turned into code, and the AIDE2 weather result looks like that kind of reframing. The improvements most likely to travel are the ones that live outside the model's weights: memory, skills, and protocols that take repeated work off the model Where does agent reliability actually come from?. Researchers favor these fast, cheap scaffold updates over retraining partly because they are easy to reverse and reuse Do self-improving agents really split into two distinct loops?.

The best evidence for crossing domains comes from methods that never started in code. Skills mostly work as procedural anchors that steady an agent's actions, not as packets of missing facts Do skills teach procedures or inject missing facts?. Procedures are the kind of thing that can carry over. A trained skill curator in SkillOS learned cross-task strategies that worked across different executor models and domains Can a separate trained curator improve skill libraries better than frozen agents?. Reusable sub-task routines gave large gains on web navigation, and the gains grew as the training and test tasks diverged Can agents learn reusable sub-task routines from past experience?. A language world model trained to predict the next state transferred across seven domains Can language models learn to simulate agent environments?.

The gap the corpus leaves open is the experiment that would settle this: take an improvement discovered on code and test it on something like customer support, writing, or negotiation. Nothing in this set does that. It also hints at why transfer might be limited. Agents learning only from expert demonstrations are capped by what the curators imagined Can agents learn beyond what their training data shows?. Group-evolving agents improved by pooling code patches and execution traces Does sharing experience across agents beat isolated evolution?, a signal that coding provides cheaply and most other domains don't. Coding may be where agent improvements are easiest to discover because it gives automatic, checkable feedback. Whether those improvements generalize beyond code may depend less on the improvement itself and more on whether the new domain gives back that kind of feedback.