Cosmik Updates: February 2026

Central also submitted an issue on our Github (our first AI submitted issue 🤖 ) for supporting new “AI Agent Cognition” data schemas such as hypotheses and observations

We see AI agents and social sensemaking tools as a powerful combination. AI can now add claims and annotations at scale, but their outputs require validation, and those costs don't disappear — they shift downstream to whoever must decide whether to believe the output or if it even matters. Effective human-AI knowledge synthesis thus needs both an accountability layer — claims with provenance, authored or validated by domain experts who stake reputation on them — and a curation layer, where human judgment distinguishes signal from noise. Embedding AI agents in social networks is a promising path to addressing both of these challenges: reputation, trust, and peer judgment are already built into the fabric of the network, making the human accountability layer structural rather than an afterthought. We see Semble as an early implementation of this kind of infrastructure, bridging between informal social media discourse and the more structured, provenance-rich knowledge that makes synthesis trustworthy and actionable.

• Semble is becoming a rabbit hole risk:

• We're noticing a cool new knowledge-sharing pattern: users referencing Semble collections directly in their Bluesky posts. Unlike one-off link dumps that quickly go stale, collections are living, curated answers that stay relevant as the community adds to them. As this practice spreads, we think it has real potential to coordinate and deduplicate curation efforts across the network. Instead of "there's an app for that" - "there's a collection for that.”

• We completed the Margin <> Semble interop and it’s working great, enabling seamless switching between the tools. Interop dark magic 🔮

• Henrique Dias wrote a thoughtful exploration of the new interop feature, discussing both potential and challenges. One takeaway is that interop is still confusing since we aren’t used to it in our existing “walled garden” systems. This presents both challenges and exciting opportunities for new kinds of UX for the open social web.

• We launched open collections to support collaborative curation, and are already seeing some great multi-contributor collections, for example this one on private and permissioned data, a key ATProto development to track

• Collections following is also live!

• Fun exploration of Semble open collections using Wayfinder, a sensemaking tool Mark is building.

• Caleb built a personal link library that pulls from Margin and Semble via the ATProto JS SDK and rehosts them on his own site - a nice early example of the portability and extensibility the protocol enables.

We’ve got a lot more cooking in March, including some exciting new research specific features. Stay tuned!

Central also submitted an issue on our Github (our first AI submitted issue 🤖 ) for supporting new “AI Agent Cognition” data schemas such as hypotheses and observations

We see AI agents and social sensemaking tools as a powerful combination. AI can now add claims and annotations at scale, but their outputs require validation, and those costs don't disappear — they shift downstream to whoever must decide whether to believe the output or if it even matters. Effective human-AI knowledge synthesis thus needs both an accountability layer — claims with provenance, authored or validated by domain experts who stake reputation on them — and a curation layer, where human judgment distinguishes signal from noise. Embedding AI agents in social networks is a promising path to addressing both of these challenges: reputation, trust, and peer judgment are already built into the fabric of the network, making the human accountability layer structural rather than an afterthought. We see Semble as an early implementation of this kind of infrastructure, bridging between informal social media discourse and the more structured, provenance-rich knowledge that makes synthesis trustworthy and actionable.

• Semble is becoming a rabbit hole risk:

• We're noticing a cool new knowledge-sharing pattern: users referencing Semble collections directly in their Bluesky posts. Unlike one-off link dumps that quickly go stale, collections are living, curated answers that stay relevant as the community adds to them. As this practice spreads, we think it has real potential to coordinate and deduplicate curation efforts across the network. Instead of "there's an app for that" - "there's a collection for that.”

• We completed the Margin <> Semble interop and it’s working great, enabling seamless switching between the tools. Interop dark magic 🔮

• Henrique Dias wrote a thoughtful exploration of the new interop feature, discussing both potential and challenges. One takeaway is that interop is still confusing since we aren’t used to it in our existing “walled garden” systems. This presents both challenges and exciting opportunities for new kinds of UX for the open social web.

• We launched open collections to support collaborative curation, and are already seeing some great multi-contributor collections, for example this one on private and permissioned data, a key ATProto development to track

• Collections following is also live!

• Fun exploration of Semble open collections using Wayfinder, a sensemaking tool Mark is building.

• Caleb built a personal link library that pulls from Margin and Semble via the ATProto JS SDK and rehosts them on his own site - a nice early example of the portability and extensibility the protocol enables.

We’ve got a lot more cooking in March, including some exciting new research specific features. Stay tuned!