Optional · weekly note
Get one useful lesson each week
One idea for building, checking or operating an AI application. Free, separate from registration, and easy to leave.
Prof Rod · Room 1
Today's lesson
Build AI systems
you can explain.
Learn to use coding agents on useful tasks—and to build, evaluate and operate AI systems. Start with work you understand, or go deeper into applications, customer integrations and systems already deployed, including the skills used in AI forward-deployed work.
Define the task → Build the workflow → Check the result
Lesson one · open to everyone
No account neededStart Lesson OneTaught by Rod Rivera — Professor of the Practice at ITAM, teaching at CDTM/TUM, after fifteen years of AI at Alibaba Cloud, Huawei, Samsung and Rasa.
Lesson One · open to everyoneRead the opening lesson before choosing a course.
Claude Code for Business3 min readLesson 1 of 12
Why This Course: The Backlog Nobody Ever Delegated
You have a reconciliation to finish, a contract folder to organize, or a recurring report to prepare. Before handing it to an agent, choose a small example whose inputs and expected result you understand. That gives you a way to tell useful work from a plausible answer.
Your experience helps you describe what matters: which records belong together, what counts as a mismatch, and which action needs approval. This course turns that understanding into a brief, a bounded attempt and a check of the result. It does not assume that a confident answer proves the task was completed.
Hall pass
Keep a lesson in your week
An optional weekly lesson to put an idea into practice. Courses are free, with registration where required; Lesson One needs no account.
The book · free to read
Build Your Always-On AI Agent From Scratch
Build a self-contained Python agent with tools, memory, messaging, permissions and recovery. Sixteen construction drafts, runnable checkpoints, and no account required.
Open the book
How we teach
An explanation. A build. A difficult question.
Understand the mechanism, build an example, then test what it can and cannot do. Bring what you know about data, software or customer work; learn the parts that are new to you.
Prof Rod
The Teacher
Explains the mechanism, draws it on the board, and works through the example. You get the reasoning as well as the result.
Mator
The Builder — visiting
Does exactly what you asked, immediately, at full speed. Which is how you find out what you actually asked for.
Quackster
The Alchemist — visiting
Asks the question you were too polite to ask. Usually the one that mattered.
The animated episodes are still being drawn. The writing, the courses and the notes are all here now.
What's in the room
The syllabus
The course register
From your first working example to systems you can evaluate
Syllabus
Live · updated weekly
IntermediateZEO ITAM · Autumn 2026
Run a measured Zero-Employee Organization (ZEO) loop against your own minutes baseline.
Serial
Subject
ITAM
Lessons
41
Time
766 min study
Open this course
766 min study41 lessons
Syllabus
Intermediate
Tools, Memory, and Multi-Agent Systems
Design tools an agent can call safely and memory it can trust past one session.
Method
Subject
AI agents
Lessons
34
Time
319 min read
Open this course
319 min read34 lessons
Syllabus
Beginner
The 12-Week Transformation Tracker, Run on Claude Code
Log whether automation is actually getting cheaper, against your own Week 1 baseline.
Outcome
Subject
Claude Code
Lessons
13
Time
91 min read
Open this course
91 min read13 lessons
The journal
Latest from the journal
Harness Engineering2026-09-2011 min read
AI Tool Error Handling: Don't Fake an Empty Result
AI tool error handling must distinguish an empty catalog from unavailable data. Use separate result types and a mutation test that catches a swallowed timeout.
Rod Rivera
Harness Engineering13 min read
AI Agent Tool Design: One Job Per Tool Boundary
AI agent tool design should expose effects and permission boundaries. Compare an overloaded inventory function with explicit read and draft tools.
Harness Engineering11 min read
AI Agent Latency Measurement: Find the Real Delay
AI agent latency measurement means tracing the whole user-visible path, not just timing the model call that feels easiest to optimize.
Harness Engineering11 min read
Agent Evaluation Variance: Did A Really Beat B?
Agent evaluation variance can make one prompt look better by chance. Design paired repeated trials, report critical failures and inspect discordant outcomes.
The corkboard
Pinned up during the week
Short notes from the practice — what broke, what it cost, what turned out to be wrong.
Measure Where the Waiting Time Actually Goes
Sep 20 · 2026The model may not be the slowest part of your AI workflow. Retrieval, tool calls and repeated validation attempts can consume most of the waiting time. For a synthetic support task, record the duration of input preparation, the model request, each tool call…
openailatency
Use Batch Work Only When the Deadline Fits
Sep 19 · 2026An overnight document-classification job and a live customer conversation have different timing needs. Choose the API workflow around the deadline. For work that can wait, evaluate the Batch API using its current completion-window and pricing documentation…
openaibatch-processing
Match Batch Results by Identifier, Not Position
Sep 18 · 2026A batch job can return results in an order different from your input list. Joining by row position risks assigning one answer to another document. Give each synthetic input a unique identifier and preserve it through the Batch API workflow. When the results…
openaibatch-processing
Measure Cache Hits Instead of Assuming Them
Sep 17 · 2026Two requests can look similar to a person while failing to reuse a cached prefix. Measure the actual usage data. Run a controlled test with a stable prompt and several different synthetic inputs. Record the API's cached-token information, total input and…
openaiprompt-caching
Put Stable Instructions Before Changing Inputs
Sep 16 · 2026If many API requests share the same instructions, keep that stable material together. Changing text at the start of every request can prevent useful prompt-prefix reuse. For a classification workflow, put the fixed task definition and examples before the…
openaiprompt-caching
Give a Tool Failure a Shape the Agent Can Use
Sep 15 · 2026A tool response that says only “something went wrong” leaves the next action to guesswork. Return a contained error that distinguishes useful recovery cases. For an order lookup, separate “not found,” “not authorized” and “temporarily unavailable” without…
openaitool-calling
Thought 1 of 0
Who teaches this
Rod Rivera
Professor of the Practice in AI at ITAM, and fifteen years of production machine learning at Alibaba Cloud, Huawei, Samsung and Rasa. Rod runs these systems on his own work first, and the classroom gets the implementation decisions, the checks and the parts that did not work.
More about RodBefore you go
Take one idea back to your work.
A free weekly lesson on building, evaluating or operating AI applications. Bring a question, try the example, and decide what transfers to your work.
Sign-up sheet
Put your name down
Free registration. Unsubscribe in one click.
