Reproduce “Every Software as an Agent: Blueprint and Case Study” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Mengwei Xu

JiT-Codegen presents a white-box software-agent architecture in which language models generate JavaScript at runtime and execute it within an application, using source code, internal state, data structures, and functions rather than only GUI actions or predefined APIs. In a 10-task prototype spanning the Electron applications MarkText and YesPlayMusic, GPT-4o completed 8 tasks, while the other tested models completed 6–7, demonstrating feasibility alongside unresolved challenges in semantic verification and security.

Key insights

  • Direct runtime access can replace lengthy GUI sequences with short, task-specific programs that operate on application state and dispatch internal actions directly.
  • Execution feedback helps correct explicit runtime errors, but it is less effective when code runs successfully without satisfying the user's intended goal; semantic task verification remains a central limitation.
  • The approach could complement existing APIs and GUI agents: generated actions may serve as on-demand APIs and later reveal capabilities worth exposing as stable interfaces.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.