Field notes, research breakdowns, and product thinking from building with AI agents every day.
Agents, information systems, and the firm, with Jack Dorsey's essay as a starting point.
Hierarchy was always an information-routing system. If agents are to participate in coordination, the real challenge is architectural: how should firms represent state, memory, and authority?
Rao & Georgeff wrote the BDI architecture in 1995. It has 3,900 citations. And it explains exactly what's broken about how AI planning tools work today.
The Belief-Desire-Intention model from 1995 is the most cited paper in multi-agent systems. Its core insight — that an agent's intentions must remain consistent with its beliefs — explains exactly what's missing from modern AI planning tools. Here's how we're building it into AgentPlanner.
Andrej Karpathy ran 700 experiments in 2 days. The real story isn't the speedup — it's the loop.
Karpathy's autoresearch experiment ran 700 ML experiments in 2 days and found an 11% training speedup. But the deeper insight isn't about model training — it's a generalizable pattern for any goal-driven agent system. Here's what it means for how AI agents should plan.
Most companies are still using AI like a smarter search engine. Here's what it looks like when you treat agents as team members.
There's a difference between using AI tools and agent-driven transformation. One is augmentation. The other is restructuring how work gets done — with agents as active participants. A phase-by-phase playbook for getting there, with AgentPlanner as the coordination layer.
Planning tools have been designed around humans for decades. What changes when agents are the primary operators?
The fundamental assumption of every project management tool is that a human is creating, assigning, and completing tasks. That assumption is breaking. When agents are the primary operators of a plan, the design principles change completely — and most existing tools aren't built for it.