SF : Continual Learning Hackathon

Spent Friday at Intercom for a one day hackathon focused on continual learning agents, with integrations and inspiration from You.com, Intercom, Lovable, and Google DeepMind. The constraint was sharp. Agents had to keep learning while running, not just call a frontier model and stop, which pushed teams toward architectures with feedback loops, persistent memory, and online evaluation rather than another wrapper around chat completions. Categories spanned AI copilots, autonomous research agents, customer support brains, real time data explorers, and creative design systems. By demo round a surprising number of teams had something genuinely usable end to end, not just a stitched UI. You.com brought search infrastructure, Intercom opened up customer support primitives, Lovable provided the rapid build surface, and DeepMind contributed integrations and inspiration on the model side. The room had the kind of energy you only get when everyone in it actually finishes things. About nine hours of building, real shipping, and several conversations that turned into longer threads later in the week. The takeaway I keep coming back to is that continual learning is the right framing for where production agents are heading. Static models with retrieval get you a long way, but the products that win in the next cycle will be the ones that improve from interaction without quietly falling apart. Thanks to the Intercom team for hosting and to the partner sponsors for showing up with real surfaces to build against.