The paper introduces Agentic Just-In-Time Software Construction (A-JIT), a paradigm where applications are built from code, a runtime harness, and an embedded AI agent. The agent continuously observes usage and execution traces to dynamically synthesize missing logic, workflows, and interfaces, much like a JIT compiler specializes machine code at runtime. This enables applications to evolve perpetually without requiring developers to pre-build all functionality.
The authors demonstrate A-JIT through examples such as a weather forecast conversion function marked with a 'code hole' that can be filled later, and a spreadsheet assistant that clarifies ambiguous user commands. By integrating synthesis into the application lifecycle, A-JIT blurs the boundaries between developer, AI agent, and end-user. The work opens a new design space for adaptive, self-evolving software systems.