A curious thing happens when you step away from the toy problems of static, one-shot bug fixing and ask an AI to actually maintain software. It turns out, the illusion shatters.
Recent research introduces SWE-CI, a benchmark built on the Continuous Integration loop to measure long-term maintainability rather than short-term party tricks. They gave these cutting-edge, supposed "AGI-adjacent" models tasks mirroring real-world repository evolution—complex requirement changes, long-term feature iterations, and the kind of architectural decisions that human engineers agonizingly debate over.
The result? The best performing AI agent could only keep 70% of the code functional without introducing a regression.
The 30% Regression Tax
Let that sink in. The absolute state-of-the-art—the models that are allegedly about to render senior engineers obsolete—break the build three times out of ten during sustained continuous integration. In a production environment, a 30% regression rate isn't "progress." It's an extinction-level event. It's the difference between a functional product and a continuous, rolling outage that bankrupts your company in SLA violations.
Continuous Deterioration
When an AI autocompletes a script, it looks like magic. When it tries to maintain a mature microservice architecture over 71 consecutive commits, it looks like a junior developer who lies on their resume and panics under pressure. The AI doesn't understand the system's invariants; it's just probabilistically guessing the next token based on how similar problems were solved in its training data.
When you task an LLM to evolve a feature, it doesn't refactor with intent. It bolsters technical debt. It piles patch upon patch, applying "fixes" that pass a narrow unit test while quietly detonating a critical assumption three modules over.
We are handing the keys of our software infrastructure to an automation tool that fundamentally lacks the cognitive friction required for long-term architectural maintenance. Letting an AI autonomously maintain your codebase isn't engineering. It's just automated vandalism at scale.