AcademyAgentic AI: from grounded answers to meta-cognitive agents
Build an agent that runs against real data for a real business and is right about it — or says, with its arithmetic showing, that it cannot be.
Most agents work in the demo and fail on a real Tuesday: a renamed column, a month with half its data loaded, two systems that disagree, a question the data cannot answer. They stall, or they answer anyway. This programme is about the second failure, because it is the expensive one. By the end you can build an agent that is right about real data, and that says so with its arithmetic when it cannot be.
Apply to join Cohort 1
Twenty sessions, three graded checkpoints, a capstone defended in front of graders. 20 seats. Fee to be announced.
Entry is by a free thirty-minute test of the basics, so you find out whether you can follow the programme before you pay.
Take the entrance testApply for a role
We hire engineers who can already do this. The test is sixty minutes: a few numbers, a few judgments, and how well you know what you know. Pass it and we talk the same week.
If you do not pass, you get a specific note on where the gap is. The programme covers those gaps; whether you take it is your decision and has no bearing on future applications.
Take the hiring testWhat you will be able to do
- Set up and solve the probability, estimation and optimisation problems that rewards, calibration and policy optimisation rest on — and recognise when the standard formula does not apply.
- Bind a business question to the data that can answer it, with provenance, and know with evidence when no data can.
- Build an autonomous agent: give it a goal, not a question; have it plan, act, observe and replan; prove with tests that it stops, asks and refuses when it should, and recovers from a failed step.
- Build a system that maintains its own structure: detects drift, versions its knowledge, repairs a bad change, reports what it could not check.
- Build an agent that holds a model of its domain, reasons over it, and revises it when sources disagree.
- Build an agent whose confidence is calibrated, that grades its own answers, and that changes strategy on measured failure — shown on held-out questions.
- Report three outcomes, never two: usable / refused / unmeasured, with the arithmetic on every refusal.
What we assume you already know
You already know how a language model generates text, what prompting, structured output, function calling and RAG are, how to write an agent loop with tools in Python, basic SQL, and probability and a gradient on paper. The entrance test checks this before you pay.
How it runs
Four weeks. Five days a week, four hours a day: eighty minutes taught, the rest at the keyboard, every lab continuing the same agent from the day before. Every Friday is a graded checkpoint on unseen data and an unseen fault.
The certificate states what you demonstrated at the three checkpoints and the defence. It does not certify attendance.