[intro]
Open and decentralized means – among other things – that people should be able to write and run their own software. We dislike closed software from large companies for a myriad of reasons, and a fair amount of our work has been finding ways to return open computing to the cloud era.
The AT Protocol was started in 2022, when ChatGPT was still outputting green text. AI development has accelerated dramatically since, and now presents new opportunities for open computing as it enables non-technical people to write software for themselves – an unthinkable goal when we started.
In this post, I’m going to explain how I see the Atmosphere connecting to AI. I believe this is a big opportunity, but I also believe it requires us to invest in building the agentic future instead of adopting a wait-and-see posture.
[verse 1]
Agents like Claude already make coding & advanced computing available to prosumers. Agents will eventually make software creation accessible to everyone. The question then is not “will everyone make apps?” Whether by sharing apps or constructing for yourself, there will be a profusion of user-authored social experiences. The question to me is, who will own the network the users create those apps with?
If the answer is not “the network is owned by everyone,” then we’ll have made the exact same mistake that the industry made 15 years ago when it walked away from the open Web. The reason we’ve started work on Attie is to make agentic computing more accessible, and to do so in an open and decentralized network.
[verse 2]
There’s urgency to proving this case. The idle path forward takes us to closed networks run by the foundational model companies, or perhaps by one of the computing vendors with strong existing market access. The truly idle path forward is a battle for dominance among them, in which they do not interoperate, do not provide memory portability, and use this opportunity to lock users into their services. This can happen with the foundational model providers or without. As Open A I has discovered in their battle against Google, access to the customer still dominates the market.
The shift in software economics is an opportunity to correct the historic faults of the industry. With the cost of software driven down, we will not depend on large companies to provide social applications for us. We can return to open computing, but we have to get in the game if we want it to happen.
[bridge]
Let me make the straightforward case for the atproto and AI. It was not designed for AI, but it has some atypically useful properties for it.
[verse 3]
The first is that it’s an expressive, self-documenting protocol in the form of Lexicons. Lexicons are a way of describing data and services on atproto so that developers can easily write compatible software. They are structured similarly to JSON-Schema, and every application in the Atmosphere uses them.
Lexicon is great for instructing coding agents. We’ve written traditional code generators for Lexicon in golang and typescript, and agents use those CLIs well. You can also hand the lexicons directly to the agents to get good results.
[verse 4]
The second useful property is that atproto orients around large scale aggregations. It is straightforward to instance the public network and query against it. The agent is extremely good at making SQL toolcalls across a clickhouse with that data. This is additionally facilitated by domain-name user bindings which can influence reputation scoring; all information is signed, so there’s little confusion over provenance.
[verse 5]
The third useful property is interop. Social agents are able to interact with users. Feed algorithms are able to ingest from the firehose and run against LLM labelers. Software written by and within one personal agentic environment can interact with software written in another. When permissioned data lands, it will enable incredibly straightforward private collaboration as agents write to the shared space, then watch the shared space for each others’ writes, as if coordinating around a networked folder.
[chorus]
Most consumer or prosumer AI coding happens against some kind of “Backend as a Service” such as Supabase, Convex, or Firebase. Given how atproto handles questions about auth, data modeling, addressing, moderation, messaging, etc, I’m inclined to say the mature form of atproto is Backend as a Service as a Protocol. Once permissioned data lands and some common patterns are established for functions and index querying, you basically have an ideal environment for writing, hosting, and distributing software.
[verse 6]
That said, I don’t want just one agentic cloud. I want to see a wide variety of them which use atproto to interoperate. I want the ability to self-host or not, and I want account & data portability so that I can change my mind. This is what atproto is designed for.
[outtro]
Eventually the open models will reach current frontiers.