FADE

• 

  Onboarding Part 1

• 

  Onboarding Part 2

• 

  Dashboard

Inspiration

We noticed that recently more and more people use AI for very mundane tasks. This excessive usage results in excess token use which causes significant costs and harms the environment. We realized that a lot of these tasks don't need AI and can be executed with scripts. That led us to FADE, the AI agent that eventually eliminates itself from the workflow, allowing for better efficiency and saving tokens.

What it does

FADE is a self-auditing AI agent that completes your task, reflects on it's own execution trace, and replaces every step that doesn't require it with a python script. At the end it will compile everything into a digest and send it to the user's preferred

How we built it

Architecture The system is built entirely on Google Cloud Platform:

Vertex AI (Gemini 2.5 Pro) — powers the agent during Execute, performs trace classification during Reflect, and generates replacement scripts during Replace Cloud Functions — hosts the generated deterministic scripts as serverless endpoints Cloud Scheduler — triggers the automated pipeline on a cron schedule matching the original workflow cadence Cloud Storage — stores execution traces, generated scripts, and deprecation reports Cloud Run— serves the deprecation report dashboard We drew out how the structure worked on a whiteboard, how the agent would split the task into multiple parts, the categories of the tasks and whether or not it would need AI to do it.

Challenges we ran into

We ran into 2 main challenges. The first was our Gemini summaries, while checking if our summaries were properly being outputted, we ran into a few difficulties. Whether that be the summary not showing up at all or only part of the summary being generated. The next challenge we faced was our website, it wasn't updating on the page when we selected a different repo on the input.

Accomplishments that we're proud of

We are proud of our idea, because it seems counterintuitive for a self deprecating AI. That is what makes our idea unique: most AI agents are designed to do more work and keep you reliant on it, but ours is designed to do the work it absolutely has to do and takes itself out of the process for everything. We are proud we got this working in a smooth matter.

What we learned

The most important thing that we learned was that most of what AI agents do in production doesn't require AI. When we instrumented the Github Weekly Digest workflow, 7 out of 8 steps were able to be automated by scripts. We believe that this ratio is consistent for many tasks of this nature. Another important thing we learned was self deprecation was pretty hard. Most teams can show an AI agent being able to do beyond its capabilities, but the smartest agent is one that knows where it's not needed.

What's next for FADE

Currently our scope is just GitHub Weekly Digests, our vision is to make FADE applicable to a variety of tasks, even the mundane ones. We believe that AI is necessary, but only when it is absolutely necessary. By spreading our scope to multiple tasks, we believe we will be able to save thousands of dollars and save tokens. Come and FADE AI with us!

Built With

• gemini2.5pro
• genai
• javascript
• python
• react18
• requests
• restapi