What they did

The authors define A-JIT applications as integrated assemblies of conventional code, a runtime harness, and an embedded AI agent. The agent observes usage and live execution traces, then uses program synthesis to specialize application behavior for a user’s current needs.

They demonstrate the model through trace-driven human–AI co-construction, including constructing missing implementations and adding capabilities during the application lifecycle rather than only before deployment.

Key findings

• The authors show how synthesis can be placed inside a running application’s lifecycle rather than confined to development time.
• A-JIT can use execution traces and user behavior to specialize software logic, workflows, and tool interfaces.
• The demonstrated design supports generating missing implementations and new capabilities while software is in use.
• The work identifies the runtime harness as the mechanism connecting observation, agent-driven synthesis, and evolving application behavior.

Why it matters

A-JIT reframes software delivery around systems that can continue adapting after deployment. If made reliable, this approach could support highly personalized applications and reduce the need to anticipate every workflow or feature before release, while extending just-in-time specialization from machine code to application-level behavior.

Caveats

The paper primarily establishes a paradigm and demonstrates its feasibility; the supplied material reports no quantitative evaluation of correctness, latency, cost, security, or maintainability. Continuously modifying live software also raises unresolved questions about verification, rollback, reproducibility, user control, and whether agent-generated changes remain safe over long deployments.