The (decentralised) AI agent stack

Patrick Mayr -

Over the past few months we’ve seen the emergence of decentralised AI agents, which differ to centralised ones in that they are built on blockchains.

For simplicity’s sake I’ve broken down the technology stack that underpins them into three core components: apps, middleware and infrastructure.

Apps

To begin, it’s worth noting that none of these applications are agents yet. They are LLMs, because their capabilities have not yet reached the point at which they can make decisions and pursue objectives independently. They are all still heavily controlled by the developers that built them. But agents is what they aspire to be so I’ll refer to them as such.

At the application layer, we’re seeing a range of different agents emerge. The medium through which users interact with these agents is primarily through social platforms like Farcaster and Twitter. However, some apps are also end destinations.

Here is a broad categorisation of the current agents as I see it:

  • Personalities (Aethernet, Luna) - agents that live on the social feed and act as if they were just another user, posting, commenting and engaging with other users. Their purpose is to provide entertainment above anything else.
  • Assistants (Gina, Bankr) - agents that can be tagged in the social feed by users to answer questions, surface data and provide other assistance.
  • Analysts (aixbt, kwantxbt) - different to personalities in that they are less controversial and more utility focused. Users don’t call on analysts as they do on assistants, but they provide useful market related insights. For example aixbt posts about trending projects and their tokens while kwantxbt is a technical analysis agent.
  • Investors (ai16z) - agents that manage capital and make investments. For example, ai16z is an investment agent that invests into memecoins and bases its investment decisions on recommendations from its community.

From an end user perspective, the the fact that agents are built on blockchains offers two interesting features. One is that agents can have a crypto wallet. This allows them to manage capital on behalf of users. The other is that they can issue a token that users can buy. These tokens represent ownership in the respective agents and can grant holders certain rights. For example, token holders of $AI16Z can make investment recommendations to ai16z.

Middleware

Frameworks like Eliza or Rig provide tools and components that help developers create and deploy agents. The reality is that the application layer for agents is still nascent and it remains very difficult to string together the different pieces of software to create them. As more agents emerge that exhibit what can be built, we’ll see more tooling emerge that makes it easier for future developers to do the same.

Launchpads like Virtuals and Clanker provide an experience through which non developers can launch agents as well. Virtuals lets users do so on its app, while Clanker lets users do so directly on a Farcaster client.

Below the frameworks sit the models themselves. These can either be open source like Llama and Stable Diffusion or closed source like GPT and Claude. Open source models allow developers to tinker with the raw models themselves and to host them on a cloud provider of their choosing. Closed source models come out of the box and are only accessible through API access. Both come with their advantages. Open source models give developers more control and flexibility whereas closed source models require are easier to get started with and require less integration work.

Infrastructure

Models are trained on data that can come from socials like Twitter and Farcaster, onchain data like previous transactions and other offchain data that can include almost anything else. Inferences are made when a user engages with the agent. The interaction between the user and agent could then result in an onchain action, such as buying or transferring a token.

For example, if I ask Gina to buy the best performing DeFi token over the past month for me, she will run an inference through querying the onchain data she was trained on to determine which token that was, and then buy it. The transaction will then be settled on the Ethereum L1.

Today, virtually all of the training and inference happens on centralised cloud hosting platforms such as AWS and Azure. There are several decentralised compute networks like Akash and Io.net that offer alternatives for developers, but their usage remains limited.

This space is evolving at lightning speed and I expect this overview to look outdated in just a short time. I’ll try to update it in the future as my thinking evolves.