Building AI Agents

What this course is, and what it refuses to be

Most agent tutorials end at the demo: a loop that calls a tool and prints something plausible. This one starts where those stop — at the parts that decide whether the thing survives contact with real users: tool schemas that reject bad input, state that does not grow without bound, failures that surface instead of hiding, and a test suite that tells you when a model upgrade broke the workflow.

We build in Python. No framework worship: you write the loop yourself first, then see what a framework buys you and what it hides.

Curriculum

Week 1 — When an agent is the right answer

• Agents, workflows and cron jobs: choosing the cheapest structure that works
• The loop: model, tools, state, stop condition
• A minimal agent in sixty lines, and every line explained
• Failure modes: loops, drift, silent errors, confident nonsense

Week 2 — Tools that do not lie

• Tool schemas: types, enums, and validation before the call
• Idempotency and retries: what happens when the network blinks
• Read tools versus write tools, and why they need different guardrails
• Human approval as a tool call, not a policy document

Week 3 — Memory and context

• What belongs in the prompt, what belongs in a database, what belongs nowhere
• Retrieval that is worth the complexity — and the cases where it is not
• Compaction: keeping a long run coherent without a growing bill
• Privacy: where user data actually ends up

Week 4 — Evaluation

• Building a small golden set from real failures
• Scoring: exact checks, rubrics, and model-as-judge with its limits
• Regression runs before every model or prompt change
• Cost and latency budgets as first-class test cases

Week 5 — Deployment and the capstone

• Running it: queues, timeouts, observability, kill switch
• Secrets, permissions, and the blast radius of a write tool
• Capstone: an agent for a workflow you own, deployed and evaluated

What you build

A working agent, not a notebook. By the end of week five you have:

1. A repository with the loop, tools and tests separated.
2. An evaluation suite that runs in under two minutes.
3. A deployment with a cost ceiling and a way to turn it off.
4. A short write-up of what it does, what it refuses to do, and how you know.

How the five weeks run

Self-paced, six hours a week on average. Each week ships code you can run. Project feedback is written, within two working days, on the capstone and on the evaluation suite.

What you need

Python 3.11 or newer, comfort with the command line, and an API key for any major model provider. About $10–20 of model spend covers the whole course. A GitHub account helps but is not required.

Who it is not for

If you have never written a function, start with a Python course first — the agent loop is not where you want to learn what a dictionary is. If you want a no-code automation tool, this is the wrong room; the course is about the code you control.