Recent thoughts
Short articles distilled from my RFCs: agent memory, intent, tool binding, evaluation, and autonomous engineering. Grounded in real systems, no hype.
Excel is the world's most-used programming language, and it's invisible to agents
Compiling a spreadsheet into a deterministic engine that AI can use but never fudge.
read →You can't unit-test a system that answers differently every time
Evaluating agents when the same input yields different, equally correct outputs.
read →A test suite can be green and still be wrong
Testing whether the artifact serves the intent, not just whether it matches the spec.
read →My agents used to crash right before finishing
A task-scoped memory that survives the context wall instead of a last-minute lossy summary.
read →Tool selection is a translation problem, not an execution one
How an agent picks the right tool from hundreds when every user speaks their own language.
read →Hallucination is the wrong word for what my coding agents do
Why I track why code exists, not just what it does, and call the real failure specification drift.
read →I wanted an engineer, not a typeahead
Why I built an autonomous engineering agent instead of another code assistant.
read →The seven ways an AI agent forgets
A pathology-first model for agent memory, drawn from real production failures, not the abstract.
read →