Self Improving Agents

Inspiration

I have been working with agentic AI / generative AI for a bit now (~1 year), and I have been generally interested in trying to auto-update AI systems to do things aligned with user feedback, so self improving agents is a good project for this.

What it does

Allows users to configure agents and their system prompts and autonomously update the agent's system prompt via chat interactions, while also allowing the user to toggle optimization off/on depending on user preferences.

How we built it

Using Effect.TS (https://effect.website/docs/), Tanstack + React + Vite, and PostgreSQL, I implemented this web application.

Challenges we ran into

The hardest part is getting all the infra in place to begin testing the agent learning process, and then testing the agent loops.

What we learned

Building agents is hard, but if you maintain a clear intent and focus, then it is tractable to get something interesting working.

What's next for Self Improving Agents

Next is building out the proper product layer (i.e. proper chats / threads within conversations, adding more settings for agent optimization, and more testing of the optimization loops)

Built With

• effect
• postgresql
• react
• vite