Bluesky has spent years making a case that social media does not have to be locked inside a few giant platforms. Introduced in closed beta at the ATmosphere conference, the company’s new Attie makes the case that your feed and interface to social media should be yours as well.



Attie is a natural-language agent built on the AT Protocol. Its promise is that a user can describe the kind of feed they want, and the system builds it. Ask for more reporting on sustainable practices, or more balanced political commentary or more neighborhood arts coverage, and Attie turns that request into a working feed without having to fiddle with obscure settings.



In a market dominated by black-box recommendation systems, Attie presents an alternative to the core assumption of the modern social web that platforms should decide what matters, and that those platforms should also determine how relevance works and which signals deserve priority in order to put posts or items on your feeds. Bluesky is arguing that AI should reflect the user’s intent and interests, not quietly reshape it for the benefit of the platform.



Feeds Built from Interests And Not Inference



Most major social networks work in roughly the same way. They watch data points, measure engagement and activities and then infer from there based on what will keep you longer and more engaged on the platform. All user’s behavioral activities are factored in such as clicks, hovering or lingering on post, follows, shares, and even scrolling behavior. All of it becomes raw material for a recommendation engine designed to keep attention in motion. That model has created some of the largest technology businesses on earth because it feeds the advertising and subscription-driven business models that reward that sort of engagement.



Still, it leaves users with very little real control. People can mute or follow accounts, maybe tweak a few ranking options, but the underlying logic stays hidden. The system keeps learning from behavior, and the company keeps control over how that behavior gets interpreted, and the algorithm controls what you can see. Attie flips that around, at least in theory.



Instead of forcing the machine to guess what a person wants, it lets the person say it. That turns preference from inferred signals into explicit instructions. That may prove more important than any one feature inside the product. The fight here is not really about feed customization. It is about who gets to define the terms of personalization in the first place.



Why The AT Protocol Changes The Stakes



If Attie were just another AI layer dropped into a closed platform, it would still be interesting. The reason it draws more attention is the foundation underneath it. Bluesky built Attie on the AT Protocol, an open, federated standard for social applications. That matters because open protocols create the possibility that identity, data, and social graphs do not have to remain trapped inside one company’s walls. In principle, a user can carry more of their social presence across services, and developers can build on shared rails instead of asking permission from a gatekeeper.





In that context, Attie is doing more than generating feeds. It is making a larger claim about how AI could fit into an open social stack. For years, personalization has been one of the strongest forms of platform lock-in. The service gets better at predicting what holds your attention and your habits become part of its moat. The recommendations improve, at least enough to keep you from leaving, and the company keeps the full benefit of that accumulated learning. Users rarely get transparency into why the feed looks the way it does, and they almost never get portability.



Attie points toward a different way of getting engagement. If personalization can be described in plain language and built on open infrastructure, then recommendation logic is much more understandable and controllable. That does not automatically solve the problem of lock-in, but it does challenge the idea that personalization has to belong to the platform alone.



The Technical Problem Is Harder Than A Demo



One of the challenges of making this work is that LLMs are not precise and people aren’t always the best at prompt engineering. They ask for contradictory things and leave out context. They change their minds and can use fuzzy, emotional and often inconsistent language to describe their interests.



A system like Attie has to turn those messy requests into feed logic that actually works. Not just once in a demo, but repeatedly for lots of users under unpredictable conditions. Large language models are good at translation, summarization, and routing tasks, but they are less reliable when nuance piles up, edge cases multiply or the request depends on subtle cultural distinctions. A feed request that sounds perfectly reasonable to a user may still produce an experience that feels skewed, brittle or overly literal once the machine executes it.



Likewise, feeds might take advantage of those inconsistencies. Any tool that lets users shape information flows in natural language will run into questions about abuse, safety, ranking and visibility. If the system builds feeds from open-ended instructions, it has to decide what kinds of requests are acceptable, which sources get trusted and how to handle outputs that technically satisfy a prompt while still producing a bad social outcome.



If Attie gains traction, the incumbent platforms might also add more AI-driven feed and customization to their platforms. But they will have a harder time copying the underlying premise. A large ad-driven platform can add a conversational interface to customize feed preferences and even layer more user controls into the product they can market as user empowerment. But that does not mean it will give up the economics of centralized ranking, proprietary models and engagement-led incentives that form the basis for their revenue.