Inspiration This project was inspired by personal headaches and dealings with what seemed like a never ending sea of open browser tabs and Desktop apps on my workstation on a daily basis. These tabs which if they were note moved to bookmarks simply remained within my browsers cache due to fear of forgetting what I thought I wanted to return to, or the endless nights of closing my computer down with several running desktop apps in the dock attempting to remind myself to start on it again in the am. Tabilize Agent Spawn has been something i have been attempting to perfect for several years, its went through various names, stages, and developments, but always with the same objective, optimize productivity.
What it does: First the program ingests all your browser tabs, browser bookmarks, and recent tab history if you so choose, along with your desktop open and running programs, to compile a spindle like animation of screenshots which allow the program to summarize, categorize, and rate the significance of each tab so that it can be placed on a destructable timer, anytime during the destructable date, user can choose to spin up a task specific agent based on the summary, category, and intention for the tab to have it completed agentically. Ideally this should provide users with a cleaned and tasks completed desktop workstation nightly.
How we built it: First Ill go back to my inspiration I saw the spinning image carousel and thought I just had to use it so shot out to Shakuro Gallery for the webgl imagery and css build. Then I utilized trusty ole javascript extension jsx to include the openai import and for the nemotron model and all other models utilized, and the rest of my tech stack was pretty simple from there just Reat JS, json files for the manifest to add to the various browsers, i built a swift mobile app version for ios, then some standard html, and css to top it off.
Challenges we ran into: The self producing Agents from each tab to complete tabs is a pain and still a work in progress, as its difficult to produce an consistent Agent based on a singular LLM to generate predictable prompts for Agents which will eventually run on various other LLMs, mostly the prompt engineering aspect of the build, everything else was a matter of semantics.
Accomplishments that we're proud of: Well at current havent actually completed the App so if completed in time for the cut off date submission and complete all the tasks required to submit that will be a reason to be proud as any.
What we learned: That employing viable deterministic AI Agents is tougher than it sounds. Sure anyone can spin up an agent with an .md skills file or through an MCP server but getting the Agents to produce verifiable valuable results is still something that seems to be a disconnect between the training of the models and the consumers expectations.
What's next for Tabilizer Agent Spawn: We will see!
Log in or sign up for Devpost to join the conversation.