Blog
What I think matters, and why.
Owning the tools around the model
The useful part of a coding-agent setup is the judgment built into it, including when to stop.
What a free service still costs
Software can be cheap to copy while the work of keeping it useful remains expensive.
What should a research scaffold decide?
A research workspace needs conventions that preserve evidence, including evidence that refuses to fit.
The tools an agent can actually use
Giving an agent a task also means giving it a workable way to act and observe the result.