Most coding agents stop learning once they are deployed. While software evolves, they remain static.

alphaXiv has recommended a new survey from researchers at Nanjing University of Science and Technology and Nanjing University (Hao Zhou, Haichuan Hu, Ye Shang, Quanjun Zhang), which redefines coding agents as systems that should continuously evolve from their own coding history.

The core argument is straightforward: compiler errors, runtime traces, code review feedback, and failed patches—all elements that are treated as "noise" in standard workflows—are exactly the learning signals that agents should leverage.


The paper maps out the evolutionary dimensions of coding agents: from memory mechanisms, tool usage, scaffolding, to model policy and multi-agent collaboration workflows. Only when all these dimensions can be dynamically adjusted based on project history can coding agents become adaptive software engineers, rather than one-off tools.

The core pain point in this field is: current coding agents typically start from scratch for every new session, resulting in discontinuous understanding of the codebase. For the same repository and the same error, they have to grope from the beginning every time. If every build, test, and review is turned into a learning opportunity, an agent's capability boundary will keep expanding over time as it is used.

One commenter on the post shared a similar take: "An agent that can actually learn from failed patches, tests, and repository history is way more useful than one that starts from zero every session." Another commenter put it even more directly: "The real unlock is when agents improve incrementally from every build, every error, every piece of review feedback."

Paper link: https://www.alphaxiv.org/abs/2608.03392