Key learnings
in AI
A few things I have come to believe from building with AI over the past few years, from inside an early Microsoft partnership to running models on my own hardware.
AI is a value chain of agents
In 2022 I was leading our partnership with Microsoft as a Copilot development partner. I spent that year in regular conversations with Microsoft leaders, exploring what extensibility of Copilot would look like and how you would orchestrate many AI tools together. It kept resolving to one idea. Instead of a single model doing everything, you build agents, each responsible for a nuanced part of a value chain. A real workflow is not one call to a model. It is dozens, and more likely hundreds, of agents working in parallel, each owning a narrow job. That is where this is all headed.
Not one model doing everything, but hundreds of agents, each owning a narrow job.
Building with AI is its own literacy
The models keep improving, and quickly. We have reached the point where knowing how to direct them, where to constrain them, and when to keep a person in the loop matters more than any single model choice. The skill is not prompting. It is system design: deciding what the machine is allowed to do, proving it is safe before it ships, and building the guardrails that let it run without supervision.
A system earns trust by what it refuses to do
The most important part of MOCO is not what it produces, it is what it will not. Every piece of work is generated under a hard constraint to a client’s verified facts, and the gates fail closed: when something cannot be proven safe to publish, it holds rather than ships. Reliability is built out of refusals, not just capabilities.
Put together, these point one way: toward people and small companies who can build with AI themselves, orchestrate their own agents, and run them on infrastructure they control.