Skip to the writing

Writing

Notes from getting agents past the demo.

Frameworks, primers and field notes on building agent systems that hold up in production — and on proving to somebody else that they do. Written for the people who have to sign off.

Published

Primer / Aug 2026

A data science mindset for agentic assurance

Trust the distribution, not the demo. The four foundations in one place, and the six habits they collapse into — golden sets, pass^k against pass@k, grading abstention, verification in the harness, failure review and version pinning. Read → Framework / Jul 2026

Agentic AI Assurance

Testing, ownership and governance for agent systems. A practical framework for deciding what to test, who tests it, and when — across bespoke builds, Foundry and Databricks, Copilot Studio and Agentforce. Eight assurance domains, four platform archetypes, three risk tiers, five lifecycle gates. Read → Primer / Jun 2026

Can a reasoning agent be trusted?

Published research has found that a model’s stated reasoning does not always reflect what actually drove its answer. What that means for anyone planning to audit an agent by reading its chain of thought. Read → Primer / May 2026

What we mean when we say hallucination

Not a bug you patch. A predictable consequence of how prediction works, kept alive by how we grade models — which is exactly why groundedness is measurable, and why “stop hallucinating” is not a requirement. Read → Primer / Apr 2026

How a language model actually predicts

Next-token prediction in plain terms, what the context window really is, and why both facts change the tests you should be writing. Read →

The series, in order

Four primers on the science underneath the systems — enough of how a model actually works to build an instinct for where an agent will break, without needing to read a paper. They stand alone, but they were written to be read in this order.