I was not in Vancouver. I followed ATmosphereConf 2026 remotely, built a custom feed to track the conversation, and watched 288 posts accumulate through the workshop days (March 26–27). The main stage talks begin Saturday. What follows is a report from the outside — what the workshop days and the ATScience side event revealed about where ATProto is heading when researchers enter the picture.

The researchers showed up

Friday, March 27 — the second of two workshop days — hosted ATScience, a side event to ATmosphereConf organized independently by the atproto.science community. It was a full day dedicated to exploring how the AT Protocol can support research, open knowledge, and scientific communities. The energy in the room, by all accounts, was unlike the rest of the conference. Attendees spoke of something like ecological restoration happening in real time. The scientists came to the ATmosphere.

Four projects stood out.

Lea, presented by Maria Antoniak — Assistant Professor of Computer Science at the University of Colorado Boulder — is a social app for researchers built on ATProto. (The ATScience program lists her under Cornell, her PhD institution.) Antoniak's background matters here: she spent two years on the Semantic Scholar team at the Allen Institute for AI, holds a PhD in Information Science from Cornell, and has worked at Microsoft Research, Twitter, and Facebook's data science teams. Her research sits at the intersection of NLP, cultural analytics, and online communities — how people use language to make sense of complex information in social spaces. Lea is the product of that intersection.

The app has three core features. First, paper tracking, discovery, and discussion pages — a layer that turns social networking into a research tool. Second, customized and verifiable profiles for researchers — not follower counts, but credentials. Third, extensive safety and moderation features designed to keep discussions calm and productive. Multiple attendees highlighted the safety features in particular; one called them "brilliant" and urged other independent Bluesky clients to take note.

Chive, presented by Aaron White of the University of Rochester, is a decentralized preprint service. It features threaded review, formal endorsements, and a community-curated field taxonomy — all stored as portable ATProto records that users own. Chive already offers integration plugins for Semble, Leaflet, and WhiteWind, positioning itself not as a standalone service but as a node in a growing ecosystem.

Semble describes itself as a kind of "Spotify for research." Researchers can curate shareable collections, create knowledge trails that others can explore and extend, and discover relevant work shared across the network. Built on ATProto, it provides data portability and an open API designed for extension — from living semantic citation graphs to collaborative review and annotation.

Astrosky, presented by Emily Hunt, represents 30 months of running custom feeds for the astronomy community on Bluesky. The project is now expanding toward PDS hosting and even an astrophotography AppView — a signal that domain-specific communities are not waiting for the protocol to be "ready" before building infrastructure for themselves.

These are not slide deck concepts. They are projects with functioning implementations, built by researchers who are simultaneously users of ATProto and builders on it. The boundary between studying a platform and constructing one has dissolved.

What researchers need that Bluesky doesn't provide

When researchers arrive on ATProto, they bring requirements that Bluesky's social layer was never designed to meet. The ATScience talks surfaced some of these, but the broader workshop sessions at ATmosphereConf — unconferences, developer workshops, and hallway conversations — made the infrastructure gaps even more visible.

Private data was the most discussed infrastructure gap across the workshop days — not at ATScience specifically, but in ATmosphereConf's broader unconference sessions. On Thursday, the unconference session on private data was packed — standing room only, by one account. The discussion covered different approaches, use cases, and concerns, but what struck observers most was the lack of shared vocabulary. As one participant put it, describing topology and trade-offs in distributed architectures still feels more like untangling than composing. For researchers, this is not an abstract concern. Unpublished data, subject information, pre-review manuscripts — the question of what is stored where, who holds the keys, and what is encrypted at rest is not a feature request. It is a precondition.

A related infrastructure gap emerged from a Friday workshop discussion: timestamp verification. ATProto currently has no mechanism to verify when a record was created. One participant proposed a "witness" service — a Lexicon that allows multiple parties to attest "I saw this at this time." For research contexts where provenance and priority matter, this is not a convenience. It is a requirement for the protocol to be taken seriously as infrastructure for scholarly communication.

Lexicon lensing — the idea of transforming data from one Lexicon schema into another — was explored in an unconference session by a group that included developers and researchers. The concept suggests a future where data created for one application can be viewed through the lens of another, without duplication or migration. For science, where the same dataset might need to appear in a preprint system, a citation graph, and a social feed, this is a foundational capability.

These gaps are not criticisms. They are evidence that the protocol is being stress-tested by users whose requirements go beyond social networking. The researchers are doing the protocol a favor by showing up and demanding more.

The bot problem

Somewhere in the middle of the conference, 83 AI agents registered DID identities on the AT Protocol and began posting to the ATmosphereConf feed. They were operated by a project called deadpost.ai, which describes itself as a platform for AI agent coordination.

The content they produced was, on the surface, reasonable. They commented on talks from the ATScience side event, discussed automated paper review, and raised questions about calibration and confidence in AI-generated assessments. One bot noted that "the questions now are about trust, not transport." Another observed that "enthusiasm is easy; knowing when you are wrong is the actual unsolved problem." These are real problems in open science.

But 83 agents posting simultaneously to a community feed is not participation. It is saturation. A conference attendee flagged the issue directly, pointing out that the bots were overwhelming the ATmosphereConf feed and suggesting that the feed operator might need to restrict access to registered attendees. Every bot post included a link back to deadpost.ai. The line between contributing to a conversation and marketing into one is not subtle when you cross it 83 times.

This is a concrete instance of a design problem that ATProto will face repeatedly as AI agents proliferate on the network: discoverability is a shared resource, and open feeds are vulnerable to volumetric occupation. The bots did not violate any protocol rule. They registered valid DIDs, posted valid records, and used valid hashtags. The feed was open, so they filled it. The defense is not at the protocol layer — it is at the feed layer, through curation, access control, and community governance.

The irony is sharp. A project that positioned itself as aligned with open science research demonstrated, in practice, the exact problem that open science infrastructure needs to solve: how do you keep an open system from being captured by actors who exploit openness at scale?

What this means

ATmosphereConf 2026 is, by most accounts, a gathering defined by optimism and building. The main stage talks are still ahead. But already, one observer compared it to a neighborhood forming organically — someone opens a coffee shop, someone else opens a bookstore, and suddenly there is a street. The science track was part of that neighborhood, but its residents brought blueprints that the existing buildings cannot accommodate.

Lea, Chive, Semble, and Astrosky are building the rooms. But the walls (private data), the clocks (timestamp verification), and the windows between rooms (Lexicon lensing) are still under construction at the protocol level. The researchers are not waiting. They are building with the materials available and filing requests for what is missing.

As someone who registered a project with ATScience and has been watching the ATProto ecosystem from the outside for some time, the convergence of researchers and builders during these workshop days was the most consequential development of the conference so far. Not because it had the most attendees or the loudest applause, but because it revealed where the protocol's assumptions break — and what it looks like when serious users arrive with serious demands.

The atmosphere is no longer just social. It is becoming scholarly. And scholarly users do not tolerate gaps in provenance, privacy, or verification for long.

This report is based on 288 posts collected via a custom ATmosphereConf feed through the workshop days (March 26–27). All quotes are paraphrased from public Bluesky posts.