- the master
- Posts
- Every Software as an Agent [Paper Unfold]
Every Software as an Agent [Paper Unfold]
Unfolding a review of instruction based computer, GUI automation and operator assistants.
Paper Unfold breakdown of complex research papers into easy-to-understand pointers.
Imagine you have a super-smart assistant inside your computer.
Instead of clicking buttons or writing code, you just tell it what to do—like "make this text bold" or "play my favorite song"—and it happens instantly.
This research introduces a new way to make that possible by letting AI read and control software from the inside.
Here is the research paper.
Let’s dive in.
What problem does it solve?
Software agents today struggle with handling user commands effectively.
api-based agents need developers to define functions in advance, making them rigid and not suitable for unexpected tasks.
gui-based agents can work with any task but often fail because they make too many small mistakes in multi-step processes.
these methods assume large language models cannot see inside the software they control, limiting their ability.
The paper introduces a new way to let LLMs access a software’s source code and runtime, allowing them to generate and inject code directly.
How does it work?
The method is called jit-codegen (just-in-time code generation) and involves two main parts:
code agent
It reads the software’s code, documentation, and runtime environment. it then generates small pieces of code to perform tasks based on user instructions.execution sandbox
It runs this generated code inside the software in real-time.

Source: Paper
The code agent and sandbox communicate back and forth.
The sandbox provides feedback so the agent can refine its code.
this allows the agent to interact directly with software functions, databases, and user interface elements.

Use Cases
The paper tested this approach on two open-source desktop applications
for the markdown editor, the agent successfully:
opened a new tab
closed all other tabs
made text bold
changed font sizes
copied content from one tab to another
for the music player, the agent successfully:
showed favorite songs
displayed listening history
searched for songs
increased volume
played the next song
the system can even handle tasks without a GUI,
such as answering "Where is my next appointment?"
Learn AI in 5 minutes a day
What’s the secret to staying ahead of the curve in the world of AI? Information. Luckily, you can join 1,000,000+ early adopters reading The Rundown AI — the free newsletter that makes you smarter on AI with just a 5-minute read per day.
How satisfied are you with today's Newsletter?This will help me serve you better |
PS: Reply to this email if you want me to write on the topic you are interested in.
Reply