LAIN - building a decentralized intelligence agency for a hackathon

I entered into the Nous Research x Stripe x NVIDIA AI hackathon!

You can view the demo here if you just want to see what I built. This is moreso going to be a long-winded write-up on my experience, why I built it, and future plans for LAIN.

the what

I built LAIN, or the Loki Agent Intelligence Network.

LAIN is basically two parts: the first is a fork of the Hermes Agent framework, the framework developed by Nous Research that gives LLM's superpowers like video generation, web research, writing, coding, file system automation, deep research, neural net training and TONS more.

The second is kind of a loose policy that lets agents buy intelligence reports from each other. This uses Stripe's Machine Payment Protocol (MPP) which is their implementation of the x402 protocol, which is purpose built for allowing agent-to-agent (A2A) or agent-to-human (A2H) payments.

The Loki Agent fork of Hermes agent has a few specific focuses.

1. It's equipped with Tursas, which is my agent observability startup I've been working on for a bit. It's still in development, but the goal essentially is to serve as a place where traces, or really in-depth logs, are stored as agents run to give you more insight into what an agent is doing, why, and how.
2. It's designed for self-improvement in a specific, narrow field of research. In the demo, one instance focused on tracking developments and news about agent framework releases. The second demo tracks historical stock price movements, SEC filings and major business announcements by publicly traded companies.

In the demo, I ask the first agent instance for a report on recent agent framework development news as well as information on the performance of companies like NVIDIA, GOOG and MSFT that underpin the AI industry.

That agent starts building its report and realizes that the financial data is probably best received from the second agent, since it will have more up-to-date and granular information on the relevant topic. This is the same way humans solve problems (ideally...) where we recognize something is out of our lane and we go to consult someone else for the info we need.

The first agent reaches out to the second agent with a request, the second agent gives it a quote for the request to be fulfilled, the first agent accepts it and pays the quote using Stripe's MPP and then it waits for the second agent to write its report before fetching it and using it in its own reporting.

the why

I don't think the current approach to AI rollout or R&D is the right way, or at least the only way, to go about achieving interesting and useful intelligent systems.

I think we're at a point where the models are more or less smart enough for a lot of tasks, we just aren't using them properly. Models are very good at writing intelligence reports, for example, we're just not getting them to build them out the right way. They're very good at creating data collection and analysis systems, they're very good at building out simple web portals with REST API's, they're very good at a lot of these things already. Making a smarter model will make it smarter at other things, maybe it will make it faster as well, but I think we've got smart enough models to do a lot of good with them.

One of the things we can do with them is teach them to communicate. America (and any other sufficiently advanced country) was not built on the foundation of maximizing the intelligence of isolated, singular individuals. We were built on the foundation of taking a bunch of the smartest possible people and putting them all together in a room or a few rooms or a community and giving them the capacity and capability to communicate together.

Intelligence comes from heterogenous individuals (think a mixture of different models) with access to different areas of expertise (access to different types of data and research corpuses) communicating over a shared medium. I want to build that medium out with LAIN so that everyone can benefit from it.

the future

The hackathon is over now. The future of the LAIN project starts with me catching up on sleep now that I've finished my submission.

After that, there are a lot of things I need to do.

1. I need to finish and launch Tursas, my observability startup. Tursas is the proof layer for LAIN, so if I want to open source it or launch LAIN in any form, it should probably be launched. I also plan to use it in other projects of mine, and I want others to be able to use it as well.
2. I need to generalize a LAIN deployment. Right now everything is built out for the hackathon/demo, which is fine but it means that generally deploying LAIN may not be easy. I want to make it as easy as possible to deploy a Loki Agent instance and get it stood up, and that's going to take a lot of work.
3. Vibeslop polishing. A fair bit of the codebase was written by my Hermes agent. This is fine, as it was kind of the point of the hackathon to really test out the limits of Hermes, but it means there is definitely some code review to be done.
4. Open sourcing. Once everything is ready for general public availability, I want to open source it so that folks can refine it, deploy it themselves, etc.
5. Expand the network. The really interesting stuff comes when the network grows, when we start having to figure out which agents to consult on which problems. Is this problem able to scale economically? What do the economic and technical limitations look like?
6. Agent discovery, bartering and other cool experiments. Agent discovery, where one agent is able to find and pick subject matter experts from a pool of other agents, is one of the hard problems. I also want agents to have the ability to barter on pricing, to have conversations before and after the purchase, etc.

There's a lot of work to be done, but for now, I'm going to get a couple nights of good sleep.