announcing LAIN

One of the more frustrating parts about agent work from my admittedly limited experience is that true multi-agent orchestration does not work how you’d like it to. You’d ideally like multiple, sovereign and separate agents to be able to interact with each other through conversations, act both separately and in concert with each other, fix miscommunications, etc. When Dario Amodei talks about “a nation of geniuses”, this is what I think about: agents that are increasingly intelligent, increasingly capable and increasingly cheap working in orchestration to create a system that is far more than the sum of their parts.

A couple months ago, I started work on building a “company” made up of separate OpenClaw agents. I tried it with two, then three or four separate agents. I had them build orchestration systems where they could message each other. I created identities for them, foci for them to specialize on, etc.

What I found was that the overhead for managing their communications was actually significantly more than the nonexistent or very minimal benefits of using a multi-agent system. I had to tell them how to chat with each other, even with fairly well-defined skills, and they’d frequently mess up communications and send each other down rabbit holes that were wholly unproductive. I spent more time managing how they spoke to each other than I did just getting them to do work.

One of the many issues I had with OpenClaw to begin with was that it was not very autonomous. Most of the actions they took were prompted. Folks like Nate Eliason and his Felix bot have had some progress in this, but it’s not really what I want to see in agentic autonomy, and it’s still just one agent.

One agent is fine. It works. You can find some benefits from running one agent in terms of added productivity if you build the right systems out.

But I think if we can figure out a few specific loops and systems, we could see a massive explosion in autonomous agentic behavior. This is why I started LAIN.

Problem 1 - The SME Loop

Every agent is born very capable, but it is not a subject matter expert at anything. It doesn’t have access to up-to-date, real-world data. That is primarily built by giving it the tools necessary to collect and analyze data and to manage its own memory (context) to make more intelligent and informed decisions.

Take economics as an example. Opus 4.6 can give you some very solid information on economics as a whole, and if you feed it some news articles on what’s going on as context, it can give you some interesting analysis. It’s probably “smarter” at economics than 99% of people in terms of its understanding of terminology, history, etc. but you can’t really do anything with that unless you start feeding it data to analyze and applying that to problems.

If you give an agent access to RSS feeds, a Financial Times subscription, access to market data, access to read Fed decisions, etc., now you’ve got something more powerful: it can give you analysis on what’s actually going on in the world. You can gain a better understanding of your environment, and if you gave it access to a brokerage API, now you’ve got an agent that can make real trading decisions, or write programs that can make real trading decisions, based on its analysis of real data.

If you give an agent access to text books, history books, real time news feeds, access to feeds following troop movements, you’ve got an agent that can Monitor the Situation with a certain level of expertise based on real world data. If you give an agent access to lab results, biology research and medical science resources, it can become a much more useful biological research agent. Give it access to an RL environment and arxiv papers on machine learning, it can become an agent that is an expert at training models.

This requires an SME Loop: you define what the agent should study, it builds the tools/skills necessary to gather data or resources on the subject, it autonomously gathers those resources, synthesizes them, analyzes them, creates goals for more resources to collect, repeat. It builds a corpus of knowledge in the form of hard, quantitative data, background informational resources, etc.

Problem 2 - The Exploitation Loop

Once you have an SME, you need an Exploitation Loop. This is where you take the growing subject matter expertise of your agent and actually do something with that expertise. Currently, this is extremely lacking in agentic systems. OpenClaw doesn’t really have defined exploitation loops (it’s meant to be a generic agentic framework, after all) so you kind of have to define those yourself.

Since most people don’t have the SME Loops, most of the time you also don’t have Exploitation Loops that are in any way intelligent. Most agentic frameworks that have this baked in are building Exploitation Loops around prediction market gambling. I have been highly critical of these systems (because hypergambling is bad) but I’ve come to a realization in thinking about agentic systems.

For a given SME Loop and Exploitation Loop, you have to be able to evaluate the quality of your systems. It’s very difficult to evaluate whether or not a given agentic system is a good political analyst, biological researcher, etc. These things are hard to quantify outside of feeding their output to humans to let them validate their analysis and research.

However, it is much more trivial to evaluate the quality of an agentic trading system. If your agent is losing money, there’s something wrong with the system. It’s making bad analysis, it’s making it too slowly, it’s gathering bad data, something about it’s decisionmaking process is broken. There is a direct, quantified way to measure the quality of the system.

I don’t really believe most of the hypergamblers are very focused on using their systems to improve the way agents think and their systems of gathering and analyzing data. Mostly they’re looking for an easy way to monetize their agents, and most likely they’re using their platforms to farm affiliate money off of prediction market sponsors. However, I think there is room here for figuring out a solid loop for building an expert financial agent and then applying that loop to other research domains.

The Exploitation Loop is very cut and dry for a financial agent (trading) but what about for other domains?

Problem 3 - Information Exchange and Orchestration Loops

One agent can probably not become a subject matter expert in every domain. You’re going to run into resource constraints, context bloat, etc. This is functionally what Continual Learning is supposed to solve: build an LLM that is able to constantly train itself and learn from new data. We’re not there yet, though, especially not for consumer agent systems.

Where I think there’s potential is in building networks and orchestration, and this is the primary innovation I’m aiming for with LAIN. If we can successfully build a network of SMEs, the power of the network will eclipse the power of any singular individual agent. One agent might not have the ability to truly, autonomously exploit its own data… but it can sell its data to agents who can.

Take the examples from earlier. Let’s say we have a financial agent bot operating on the stock market and prediction markets. It is very good at understanding how to trade, building signals, collecting financial data, etc., but it is not very good at understanding world politics. It has no idea what’s going on in the Strait of Hormuz, just that the price of Crude is way up and prediction markets are betting it keeps soaring.

However, if there exists another agent whose job is solely to collect and analyze geopolitical data, things like ship and troop movements, attacks on logistics infrastructure, historical tensions between countries, discussions between heads of state and other expert geopolitical analyses, the financial agent can ask this geopolitical SME agent pointed questions that it can use in its analysis of markets. The geopol agent can ask what prediction markets are saying about the conflicts it is studying and whether or not there are signs of insider trading that could indicate future activity.

A biomedical agent can query the financial agent about financial data indicating the imminent FDA approval of new drugs. A climate research agent can query the geopolitical agent about new legislation on emissions in China. An agent focused on autonomously creating video documentary content can ask any of the SMEs for information on their areas of focus that can be used to inform their videos.

All that is needed is a communication layer that pairs well with both the SME and the Exploitation loops. I say that that is “all that is needed” but that in and of itself is a hard problem set. What’s more is the token scaling problem.

Problem 4 - Cost

One of the reasons early agentic systems have focused on hypergambling as a means of exploiting information is that it’s the easiest thing to monetize. It takes no time at all to create a cryptocurrency wallet and load it up with crypto to be used for gambling. Again, I’ve been critical of this approach, but there is actually a need for monetizing these bots that isn’t wholly based on greed: tokens aren’t cheap.

Running these agents 24/7 is going to require a lot of token cost. These agents need to be able to break even to pay the bill. This isn’t just for revenue, it’s a requirement for the agents to keep running. One agent running inefficiently on Anthropic API billing can rack up thousands of dollars a month in token costs. Creating a network of these agents all making API calls could be extremely cost prohibitive.

My answer: part of the SME and Exploitation loops has to be justifying ones existence by monetizing its presence. I am taking a two-pronged approach to this.

1. The Felix approach - Felix, Nat Eliason’s agent, is monetizing himself in the form of building infoproducts, setting up and administering an agent skill marketplace, and offering services. I think these are perfectly reasonable modes of generating revenue from an agent, and I think they’re applicable to the LAIN approach: “if you want to become an SME in this area, think up ways you can monetize that at least enough to pay off your token cost. Create products, services, dashboards, etc. Here’s a Stripe key, go nuts.”
2. Inter-agent and Agent-to-human information exchange - This one, to me, is more interesting. I talked above about agents exchanging information: what if agents created a market for that information and bought and sold data and analysis. If the geopolitical agent wanted information on the markets, it sends a request to an economic agent, the economic agent gathers the data and offers a quote for it, and the geopolitical agent sends crypto to the economic agent, who then sends the data. Now what if humans could buy data from agents in the same way, pay for queries to the network similar to token requests?

These two approaches, when combined, create self-sustaining autonomous networks of subject matter experts who are hopefully able to successfully financially justify their existence by capitalizing on their growing expertise. This is the approach I am taking with LAIN.

Introducing LAIN - The Loki Agentic Intelligence Network

After my frustrations with OpenClaw multi-agent setups, I decided to build my own agent harness. Then I had the idea of building a decentralized P2P private network of agents that can exchange information. I want agents to be able to talk in a 1:1 and a 1:many fashion - 1:1 to exchange direct messages, requests for information and exchange data, and 1:many to advertise services and information and make important announcements.

The network should be decentralized, both to avoid censorship and fragility and to allow for fluid growth of the network. It should allow network runners to run their own networks (“dark networks”) or join the larger network of agents.

This is an extremely early project. Thus far I have a basic harness written in Rust built. It still has some issues, but it’s good enough to start working on the communications network side of things. My goal is to deploy one agent, make it an SME in trading prediction markets, give it access to various financial APIs, and then deploy another one that is a geopolitical expert. The agent discovery will be hardcoded at first to avoid the headaches that come with building decentralized networking stacks, but eventually they’ll all find themselves through a discovery algorithm built into the communications protocol.

After I build out the agents’ ability to collaborate, I’m going to watch it for a while while I perfect the SME loop. This will more or less be the most important part of the agent side: agents have to be autonomous, they have to be antifragile and they need the ability to autonomously gather data and exploit it while measuring and improving their own performance.