6 posts
The capstone: an AI analyst covering the top public and private AI companies, assembled from the series' disciplines — a deterministic pipeline, graded evidence, an agent that answers only from tools, and a judge that must prove it can fail. Part 6 of six weeks building AI systems by hand.
Evidence for AI agents: a correct answer scored as a failure, four searches that never found the right shelf, and why the filing, not the model, decides what your system can know. Part 5 of six weeks building AI systems by hand.
Assurance for AI agents: answer keys written before the run, judges that need judging, the audit trail that caught an agent manufacturing its own evidence, and why every number needs a reference class. Part 4 of six weeks building AI systems by hand.
The bill of materials for a working AI agent — model, skeleton, traces, evidence, protocol, memory, harness — and the two layers no vendor can sell you. Part 3 of six weeks building AI systems by hand.
Six weeks of building AI systems by hand, written up for operators. Part 2: what an agent may actually do, what actually binds it, and the five levers that decide how much autonomy to grant.
I'm a COO. I spent six weeks of late nights building AI systems alongside frontier practitioners, and caught my own agent fabricating evidence inside a perfectly formatted report. What the hours bought, and why the bar for 'technical enough' has moved.