Weekly writing on AI strategy, architecture, and the real challenges of making AI work at enterprise scale.
All views expressed here are my own and do not represent the views of my employer.
The most dangerous thing you can do with an AI agent is trust it too quickly. Here's the case for graduated autonomy - and why the rules for gaining and losing trust should be deliberately asymmetric.
Most multi-agent systems treat disagreement as a bug. The best ones treat it as signal. Here's why deliberately biased agents produce better collective decisions.
The orchestrator model is hitting its ceiling. Here are the architectural patterns emerging to replace it — and why financial markets might be the proving ground.
MCP handles tools. A2A handles discovery. Nobody has solved how agents actually agree on what to do. Here's what that protocol would look like.
Gartner predicts 40% of agentic AI projects will be cancelled by 2027. From what I'm seeing on the ground, that number might be generous.
When every team is building agents in silos with different frameworks, you don't have an AI strategy. You have technical debt with a chatbot interface.
Everyone wants to talk about models. Nobody wants to talk about the data feeding them. Twenty years in, and this hasn't changed.