We investigate how a coding agent learns across a sequence of abstract reasoning tasks. The agent runs on a frozen foundation model within a fixed harness and acts by writing and executing Python and shell scripts. Apart from the artifacts it writes, it retains no state between turns, so every thought—any belief, rule, or plan committed to a file—leaves a trace.
We let the agent play ARC-AGI-3, a suite of interactive reasoning games that provide no instructions. Each game consists of multiple levels, and a strategy that clears one level can fail on the next, making every new level effectively a new task. The agent records what it learns as Python scripts and textual notes, while the harness keeps a complete log of every action and observation.
Our contribution is a measurement protocol that traces each thought through these files, from the task where it forms to the task where it is corrected or abandoned. We applied this to seven evaluation runs using three backbones from two model families. Scripts written for one task are almost never invoked again in later tasks (only 33 of 630 references cross a task boundary) because most scripts embed the state of the current level. Instead, the agent rewrites its knowledge into new scripts, preserving general rules while discarding level‑specific details, and abandons 74% of the scripts it wrote before a boundary.
The notes, which only the model reads, are never revised: the agent appends without removing earlier claims, and the contradictions that accumulate are settled against the log. Because the log preserves everything, the agent forgets selectively rather than catastrophically. The most costly error is a hard‑coded value carried into a task where it no longer holds. These findings come from the files the agent wrote, without accessing the model itself, providing a white‑box analysis of how a coding agent continually learns.
Review