A new paper describes software that keeps building itself after you've already installed it.
The proposal, called A-JIT (Agentic Just-In-Time Software Construction), was posted to arXiv and describes an application bundled with a runtime harness and an embedded AI agent. That agent watches live usage and execution traces, then writes new code, workflows, or tool integrations to fill gaps it notices, without a developer shipping an update first. The authors frame it as an extension of just-in-time compilation: instead of specializing machine code to a hot execution path, the system specializes the application's own logic to how a specific user actually works. It's a research proposal, not a shipping product; the paper lays out the architecture and design space rather than benchmarking a working system.
If it pans out, it's a genuine break from how software has always worked: built once, shipped as a fixed artifact, patched on someone else's schedule. A-JIT imagines an app that notices you keep exporting to a format it doesn't support, then builds that exporter itself, mid-session, using your behavior as the training signal.
That's also the kind of idea that's easy to sketch in a paper and hard to make safe in practice. Code that writes and ships its own features raises obvious questions about testing, security review, and who's accountable when the agent gets it wrong, questions this abstract doesn't attempt to answer.