Jac-Browser

Inspiration

LLMs can be extremely helpful for learning unfamiliar tools, however it's often difficult to give models a complete understanding of what you're working on. Personally, I found working with Salesforce to be an arduous ordeal, especially with custom plugins adding complexity. In between sending screenshots to ChatGPT to elaborate on the unique interface, the inspiration for Jac Browser was born.

What it does

It's difficult to properly prompt autonomous agents only through text. Whether you're detailing a complicated process, or a non-techie struggling to express yourself, Jac Browser helps get your point across. Our browser enables users to demonstrate their workflows to provide more context and deeper insight to agents. Our motto is to show and tell.

How we built it

We built the project around Electron as the browser engine and Jac as the core language for the browser logic and AI pipeline. We structured the system into three layers: an observation layer that records what the user does and what the page looks like, a compilation layer that turns those raw traces into structured workflow steps, and an execution layer that can dry-run and plan those workflows. We also built a Layer Dev panel directly into the browser so we could test the pipeline end to end while developing.

Challenges we ran into

The biggest challenge we ran into was complexity. Our idea required a lot of moving parts working together, including the runtime, helper orchestration, browser integration, and reliable execution flow. Even when individual pieces worked, getting the full system to behave consistently across the entire stack was a much larger task than we originally expected.

Accomplishments that we're proud of

We are most proud of building the foundation for a genuinely complex system in a short amount of time. Even though we did not fully implement every part, we established the core architecture, got the main pieces connected, and proved that the concept is technically possible. We are also proud that we pushed through the complexity instead of simplifying the idea too much, which gave us a much stronger base to build on after the hackathon.

What we learned

We learned that a big part of building agentic products is not just generating output, but making the system trustworthy and debuggable. A helper that sometimes works is not enough. You need clean runtime behavior, understandable states, and clear failure modes so users know what happened and why.

What's next for Jac-Browser

In the future we'd like to increase the scope of the observation layer. This would give Agents a more holistic context of user behavior and facilitate the ubiquitous integration of AI into the browser.

Built With

• jac