01
Decide whether it should be an agent at all
Separate the cases where model judgment earns its cost from the ones a deterministic workflow handles better, and say why in one paragraph a reviewer accepts.
COURSE 002 · 0.1.0 · LIVE
Agentic AI System Design: From Prompt to Production
Build, evaluate, secure, scale, and explain production-grade AI agent systems.
FOR AI Engineer candidates · Software engineers · Forward-deployed engineers · Solutions architects
- Modules
- 17
- Reading and build time
- 35-45 h
- Practice scenarios
- 12
- Quiz questions
- 51
- Price
- Free, no account
THE THESIS
A model is not a system. The system is what you are being paid to build.
What a demo proves
That a model, given a clean prompt and a cooperative user, can produce a plausible answer once. Demos run on the happy path, with one tenant, no adversary, no crash, no approval queue, and no bill.
What production demands
That the surrounding software owns authority, persistence, identity, and recovery: the model proposes, the application authorizes. Everything hard about agentic systems lives in that sentence, and this course is about the second half of it.
The course teaches one method, RUNTIME, and applies it to every design: Requirements and risk, User journey and authority, Nodes and topology, Tools and context, Integrity and identity, Memory and durable execution, Evals and economics. Read the method.
AFTER THIS COURSE
What you will be able to do
02
Draw the authority boundary
State what the model may propose and what only the application, a policy engine, or an authenticated human may authorize. Autonomy stops being an adjective and becomes a matrix.
03
Write tool contracts that survive production
Typed inputs, caller identity, scopes, side effects, reversibility, timeouts, retry class, idempotency, and reconciliation for the acknowledgment gap.
04
Engineer context instead of stuffing prompts
Rank sources by authority and freshness, keep untrusted content on the far side of a boundary, budget tokens, and carry provenance to every claim.
05
Make a run survive crashes, deploys, and long waits
Durable execution, event history, replay-safe control code, signals and timers, and versioning for work that is still in flight when you ship.
06
Evaluate the trajectory, not just the final text
Golden and adversarial cases, component and trajectory metrics, model judges with calibration, and a release gate that actually blocks a bad version.
07
Operate it
SLIs and SLOs for agentic work, traces that cross asynchronous steps, cost and latency budgets, backpressure, and a runbook someone can follow at 3am.
08
Defend the design in 45 minutes
Run the same method under interview pressure, cover the eight areas in order, and tie two tradeoffs back to the requirements.
THE SYSTEM YOU BUILD
Atlas: Counterparty Due-Diligence Agent
One system runs through all seventeen modules. Atlas ingests a counterparty due-diligence request, researches it across trusted and untrusted sources, produces a cited risk packet, and routes material decisions to a human. It is deliberately unglamorous and deliberately hard: multi-tenant, long-running, regulated, and full of actions you cannot take back.
- Every module adds one production concern to the same design.
- Every failure injection breaks Atlas specifically, not a toy.
- The capstone is Atlas, complete, with acceptance tests and a design defense.
YOUR PROGRESS
0 of 17 modules done
Saved in this browser only. No account, no server.
CURRICULUM
17 modules, 40 h 30 min of directed work
Grouped by track. Each module is self-contained enough to read alone and sequenced so the Atlas build compounds. Prerequisites are listed on every lesson.
HOW EACH MODULE WORKS
Learn, ship, run, design
LEARN
Learn the principle
Each module opens with the production problem, then the mechanism, then the tradeoffs. Concepts are grounded in primary sources, not vendor marketing.
SHIP
Ship the Atlas milestone
Every module adds one part to the same system. By module 16 you have a complete design and an implementation plan, not seventeen disconnected demos.
RUN
Run the failure injection
Each module names a specific way the design breaks and asks you to trigger it. A design that has never failed on purpose has not been tested.
DESIGN
Answer the interview drill
A prompt that forces the module's idea into a live design conversation, plus at least one tradeoff you can defend.
Then a quiz with explanations, and the primary sources for every implementation-sensitive claim in the module. Quiz scores and module completion are stored in your browser. There is no account and nothing is uploaded.
PREREQUISITES
What you need first
- You can read and write code in at least one language, and you have shipped something that other people used.
- You understand HTTP APIs, a database, and what a queue is for.
- You have called an LLM API and seen it do something useful and something wrong.
- You do not need to have built an agent before. You do not need a specific framework.
TIME AND COMPLETION
What finishing means
35-45 hours end to end for the full build, or roughly 40 h 30 min of reading and directed design if you skip the implementation. Most people take four to eight weeks part time.
A module is complete when you can answer its check questions, pass its quiz, and describe its Atlas milestone. The course is complete when you can defend the whole Atlas design in a 45-minute interview, which is what the practice scenarios and the rubric are for.
SOURCING STANDARD
Every implementation-sensitive claim maps to an official source
Framework behavior, protocol semantics, and provider limits change. This course cites primary documentation for anything an engineer would act on, and the source map lists them per module so you can check the current version before you build. Where the docs disagree with a vendor blog post, the docs win.
The AI Runtime
This course is part of AIR, a publication and course library about building AI systems that survive contact with production. Course 001 covers engineering long-running agents, and the library collects the patterns and case studies both courses draw on.