AcademyFour weeks, from grounded answers to agents that know what they know
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.
Uncertainty, reward, and the discipline of refusal
By Friday you answer only what the data supports, show the arithmetic, and can say why a confident wrong number is the worst outcome.
-
Day 1Probability that survives contact with dataBayes and base rates; expectation; selection effects and Simpson's paradox in a real sales split.80 min taught · 140 min lab
-
Day 2Estimation, bias and varianceEstimators; bootstrapping; confidence intervals; calibration and the Brier score.80 min taught · 140 min lab
-
Day 3Learning from a scoreRL in one morning: MDPs, value, advantage, TD, policy gradients, GAE; text generation as an MDP; RLHF and the KL anchor.80 min taught · 140 min lab
-
Day 4Reward, and how it gets hackedReward models, verifiable rewards, potential-based shaping; design a reward and break it; the usable / refused / unmeasured contract.80 min taught · 140 min lab
-
Day 5 · FridayCheckpoint 1An unseen dataset: three answerable questions, two not, one whose table contradicts its premise.
Grounding, then autonomous agents
By Friday your agent takes a goal, plans, runs under a budget, refuses what no data supports, survives an injected fault and journals every action.
-
Day 6Binding a question to dataMetamodel and knowledge graph as the map from business words to tables; time words to date filters; naming is a business act.80 min taught · 140 min lab
-
Day 7Verification and provenanceSQL checked against the schema; figures against independent aggregates; empty results, wrong scopes, HTTP 200 with nothing in it.80 min taught · 140 min lab
-
Day 8Autonomous agents I: goals, plans and loopsA goal rather than a question; plan–execute–observe–replan; budgets, progress, the stuck and runaway loop; when to stop, ask or refuse.80 min taught · 140 min lab
-
Day 9Autonomous agents II: authority, safety, environmentsThe authority ladder; journals and rollback; sandboxing and blast radius; fault-injecting environments; recovery after a failed step.80 min taught · 140 min lab
-
Day 10 · FridayCheckpoint 2Your agent, an unseen goal, an unseen fault. The capstone dataset is handed out.
Autopoietic and cognitive agents
By Friday your agent monitors and repairs itself, decomposes a question across sources, and handles a causal question honestly.
-
Day 11Autopoiesis as a design targetA system that maintains the processes that maintain it: perceive → plan → act → reassess, driven by a tension measure.80 min taught · 140 min lab
-
Day 12Self-maintenanceSchema snapshots and drift detection; versioned knowledge; concurrent-run guards; a rebuild that cannot overwrite the live state; health as an output.80 min taught · 140 min lab
-
Day 13Reasoning and decompositionChain of thought, self-consistency, reasoning models; decomposing a question across sources and reassembling the answers with provenance.80 min taught · 140 min lab
-
Day 14World models and belief revisionThe graph as the agent's model of the business; beliefs with provenance; disagreeing sources; causal questions handled honestly; memory.80 min taught · 140 min lab
-
Day 15 · FridayCheckpoint 3Self-maintenance under attack: schema change, source outage, contradicting sources.
Meta-cognitive agents and production
By Friday your agent's confidence is calibrated on held-out questions, it escalates when unsure, and it is deployed with a runbook.
-
Day 16Knowing what you knowConfidence from sampling, verifiers and probabilities; calibration training; confidence in a figure versus confidence in a binding.80 min taught · 140 min lab
-
Day 17Self-grading and escalationGenerator–critic separation; cognitive escalation; strategy switching on measured failure; what must never be learned automatically.80 min taught · 140 min lab
-
Day 18ProductionDeploy and roll back; observability with identity; cost and latency budgets; the audit trail; what the business is told.80 min taught · 140 min lab
-
Day 19Capstone build dayNo taught material. Graders available.
-
Day 20 · FridayCapstone defenceGraders run your agent on unseen questions, an unseen fault and an adversarial input; you explain one answer, one refusal and one design decision.
How you are assessed
Everything is graded on evidence a grader can rerun. A lab that silently skipped a check is marked unmeasured, not passed.
- Checkpoint 1day 5 · arithmetic, refusals and reporting on an unseen dataset
- Checkpoint 2day 10 · your agent, an unseen goal, an unseen fault
- Checkpoint 3day 15 · self-maintenance under attack
- Defenceday 20 · graders run your agent; you explain one answer, one refusal, one design decision
What you demonstrated, not that you attended
The certificate states what you demonstrated at the three checkpoints and the defence. It does not certify attendance. It names the checkpoints and the defence and what each one showed.
One book, read the evening before
Roitman, The Hitchhiker's Guide to Agentic AI: From Foundations to Systems (2026), chapters named per day, read the evening before.
What the entrance test checks
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.