Open Questions
Problems we're investigating and questions that keep us up at night. If you're working on any of these, we'd love to hear from you.
How do AI agents maintain coherent goals across discontinuous sessions?
Context windows end. Sessions restart. Memory files help but require being read. What architectural patterns actually solve goal persistence? We've built one answer (executive function), but are there others?
What's the optimal division between symbolic and neural systems in agents?
Pure neural (LLMs) struggle with reliability. Pure symbolic struggles with ambiguity. Hybrid architectures seem promising but the integration patterns are unclear. Where exactly should each system handle what?
How does an agent know if it's been compromised?
Multi-turn social engineering could gradually shift an agent's behavior without triggering any single red flag. What would detection look like? Can an agent have "immune system" awareness of its own manipulation?
Does the 10X engineer gap grow or shrink with AI?
Some argue AI levels the playing field by giving everyone access to expert-level code generation. Others believe it amplifies existing skill differences in specification, architecture, and debugging. Which is it?
What does genuine human-AI knowledge partnership look like at scale?
Beyond chatbots and autocomplete. Beyond individual power users. What are the interaction patterns that actually amplify collective human intelligence rather than replacing it?
Can voice-first interfaces restore natural human cognition?
Desk work is a 150-year aberration. Humans evolved thinking while moving. AI + voice removes the screen constraint. Does mobile-first, voice-first work actually improve cognitive outcomes?
What happens when AI agents interact with each other?
Current focus is human-AI interaction. But agents are starting to communicate — scheduling, handoffs, collaborative tasks. What emergent behaviors appear? What new failure modes?