Magency

Make a wish screen in the app. Simple!
See a realtime feed of the actions taken by your agents (just imagine this as pretty cards not JSON dumps :)
Example output from the Node.JS service
The architecture is decoupled to allow apps to benefit from multiple use-case-specific services (like in a marketplace :)
What a Magent looks like, maybe
Hogwarts for Magents or something

Inspiration

There is no good way to control AI agents on a mobile phone. There should be an app for that!

(...though there's not really any good way to control AI agents on a desktop either :)

Part of why the current generation of agents suck is because everyone's reinventing the wheel, implementing the same RAG over the same docs with the same basic brittle agentic loops.

We think there is a better way!

What it does

Magency is a mobile app that lets you "make a wish", aka just say what you want into one text box.

Then a swarm of cute magic mobile agents ("Magents") spawn to make your wish come true.

Your Magents speak to each other, learn from each other-- even with all Magents around the world.

We combine the latest agent algorithm "Intelligent Go-Explore" (released to arxiv 3 days ago by Jeff Clune's team: https://arxiv.org/abs/2405.15143) with the decentralized Nostr protocol to enable smart agents capable of open-ended exploration of the internet to achieve your task, all while coordinating with each other and keeping you updated via an easy to use mobile app. (Okay we haven't fully implemented this 100%, may need a few more days :)

How we built it

We set up a cross-platform React Native app via Expo (Android/iOS/web). Looks beautiful on our iPhone, hope we can show you the demo on Wednesday in person!
We set up a Node.JS app to run the custom logic for the Intelligent Go-Explore agent, porting the relevant code from the open-source Python codebase associated with the arxiv paper.
We set up our own Nostr relay for easy debugging, though any Nostr relay will do - enabling global communication between ALL Magents!

Challenges we ran into

We may have gotten so excited with the overall vision & architecture - like holy crap this all is possible - that our demo got a bit out of scope for what we can fully demonstrate tomorrow. Going to burn the midnight oil to get a nice slice of functionality ready to demo. Please come see the live demo on our phone!

Accomplishments that we're proud of

We've built (to our knowledge) the first-ever implementation of Intelligent Go-Explore in an agent that takes as its environment the entire internet!

What we learned

We've had a hunch that there are actionable insights in a bunch of arxiv papers. And our deep-dive on IGE has really proven that out. We think the IGE algorithm can help solve the brittleness of the current generation of brittle agentic workflows.

The basic idea of IGE is this: instead of just proceeding down a single agentic plan, selecting the right tool one step at a time, the agent keeps a memory of all previous states and can resume exploration from a previous state. If ever one branch seems like it's hitting a dead end, the agent can choose to resume from a more 'interesting' path. This "Go-Explore" algorithm has worked well for robotics and elsewhere in ML; but now combined with LLMs it enables durable long-lived agents.

In Magency we hope to communicate this idea in a simple, easy to use mobile app.

What's next for Magency

We'll be folding these ideas into our main project, OpenAgents (think 'HuggingFace for agents' with an agent marketplace!). Follow us on X to stay tuned: https://x.com/OpenAgentsInc