Experiments with agentic AI, run and documented — including what didn't work.
An evidence engine for a claim too big to assert without one
What every experiment in this Lab must answer for
Six steps to find out what an agent can really do — and what happens after
The craft of turning a piece of work into an agent that performs it
Twenty cells, four practices, and the state of the evidence
What the Lab tests first, and what is expected to break
An editorial operation delegated, corrected while live, and still running
Isolating the act and recording the predictions
Six scars from building while live
The numbers behind the gate
Four predictions tested, four verdicts placed
The discipline of correcting while live
What transfers when the architecture meets a new domain
New notes are added as experiments conclude.