AssayAssay

assay /ˈaseɪ/ — a test of the quality or composition of something.

Learn to build agents
you can measure, control, and trust.

10 free, hands-on courses for developers and technical teams who can call an API and want to get from agent demos to systems they can stand behind.

By the end you will be able to build an agent, evaluate it against a dataset, diagnose its failures, put verification and guardrails around it, and decide when it is safe to ship.

Start the free courseEvaluate your own agent

No signup. The first lesson takes about 8 minutes. Labs run on a free NVIDIA Build API key — no credit card, nothing to install.

Ticket triage · accuracy by category

5 models · 50 cases

shippingbillingreturnsaccountescalate
llama-3.3-70b0.960.940.700.920.90
nemotron-super-49b0.940.920.680.900.88
gpt-oss-20b0.910.870.630.860.82
nemotron-nano-9b0.900.860.620.840.80
llama-3.1-8b0.880.840.580.820.78

Every model scores lowest on returns. When five models from three labs share a weakness, the cause is usually the labels, not the models — lesson 04 shows how to tell.

01Foundations

Measure an agent, and build the basic loop everything else wraps around.

01results table

Evaluating AI agents

Measure whether an agent works, using datasets, graders and judges you build yourself.

11 lessons · 2.1 hours

02iteration trace

Agent loops

The inner loop: agents that call tools, know when to stop, and refuse the actions they should not take.

8 lessons · 1.6 hours

02Building agents

The craft: make the loop robust, govern it, give it structure and shape, and ground it in facts.

03maker and checker

Loop engineering

The loops that wrap the agent loop: verification, scheduling, and the loop that changes the harness itself.

6 lessons · 1.3 hours

04checked rules

Harness engineering

Everything that governs an agent that is not the agent: guides, sensors, preconditions, and the dataset that remembers what went wrong.

8 lessons · 1.7 hours

05execution path

Graph engineering

Make an agent's control flow explicit, so the parts that must not vary are structure rather than instruction.

8 lessons · 1.6 hours

06pattern comparison

Agent patterns

The five workflow patterns and the one agent, each run against the same task so you can see what every extra call actually buys.

8 lessons · 1.6 hours

07similarity scores

Retrieval

Give a model facts it never learnt, from documents you control — and know when it found nothing.

8 lessons · 1.7 hours

03Production

Real domains and real constraints: apply it, run it responsibly, and hold up under attack.

08checked rules

Agents for industries

Six real processes, and how a domain's rules decide an agent's shape, its guardrails and its tests.

8 lessons · 1.7 hours

09one governed agent, instrumented six ways

Agent operations

Run agents responsibly in production: risk tiers, structural limits, human checkpoints, release gates, monitoring and honest factsheets.

9 lessons · 1.9 hours

10one target agent, attacked six ways

Agent security

Attack one agent to learn to defend it: prompt injection direct and indirect, exfiltration, tenant isolation, least privilege, supply chain, and adversarial testing.

9 lessons · 2 hours

The capstone

One agent, built the whole way

A runnable repository that threads all ten courses into one support agent — evaluation baseline, tiered autonomy, a release gate, monitoring, and a factsheet. Clone it, run the tests with no key, read the deployment checklist.

Open the capstone →

How they fit together

Ten courses in three tracks, and the order is the argument. Each is easier once the one before it is done, though nothing stops you starting wherever your problem is.

Foundations

Measure an agent, and build the basic loop everything else wraps around.

01Evaluating AI agents comes first because everything else assumes you can tell whether a change made things better.
02Agent loops is the inner loop: an agent that calls tools, stops sensibly, and refuses the actions it should not take.

Building agents

The craft: make the loop robust, govern it, give it structure and shape, and ground it in facts.

03Loop engineering wraps three more loops around that one — a checker, a trigger, and a loop that changes the harness itself.
04Harness engineering is the control system. Loops are about repetition over time; the harness is about governance at every step, and it is the part that outlives any model inside it.
05Graph engineering is what you reach for when a loop stops being controllable, making the fixed parts structure rather than instruction.
06Agent patterns names the shapes. Most of what the earlier courses built is one of these patterns, and this is where you learn to pick the cheapest one that fixes the failure you actually have.
07Retrieval supplies the facts. Retrieval is one tool an agent can call, and the failure modes are its own: what was chunked, what was found, and whether the answer used it.

Production

Real domains and real constraints: apply it, run it responsibly, and hold up under attack.

08Agents for industries applies all of it to processes where the constraints are handed to you: an audit trail, a clinical boundary, a liability line.
09Agent operations The operations layer around everything else — after the agent works, this is how an organisation runs it without regretting it.
10Agent security The adversary's view of everything else — put an agent under attack, and learn the structural controls that hold when the model is fooled.

Go deeper on specific problems

One hard question each, answered properly, with a single substantial lab. Take them whenever the question becomes yours.

What is your score actually worth?

Confidence intervals for evals: why a 50-case dataset cannot see a four-point improvement, and how many cases you would need.

18 min

What should an agent emit?

OpenTelemetry for agents: the GenAI semantic conventions, what to put in a span, what to alert on, and which attributes you should think twice before recording.

22 min

What is a skill, and when does it beat a prompt?

Skills as codified, versioned, on-demand knowledge: how they differ from prompts and tools, what belongs in one, and what loading one changes — measured.

18 min

How it was made

Written alongside the code it teaches. Every lab was run against real models while its lesson was written; where a run exposed a bug in the site itself, the lesson says so rather than hiding it.

Rests on

Anthropic’s Building Effective Agents, the OpenTelemetry GenAI conventions, and Kief Morris on humans and agents. Sources are linked where a lesson uses them.

Open and current

The whole site, labs included, is on GitHub. Last reviewed August 2026; the model list is re-probed against the live API.