Production AI Agents for FDEs

Production AI Agents for FDEs is a self-paced Python course for forward deployed engineers and software developers learning to build agent systems. It starts with basic Python and develops intermediate skills in evidence retrieval, tool permissions, human approval, crash recovery and agent evaluation.

A support agent gives a convincing answer. Then the policy changes, the reviewer approves an old proposal, or a replacement request succeeds just before its response disappears. What should the system do next, and how would you prove it did the right thing?

You will answer those questions by building one system for Calder Equipment, a fictional industrial-equipment company. You begin with a small support baseline, then add applicable policy evidence, bounded model decisions, scoped tools, separate approval and recoverable actions, audit demonstrations and training signals, and finish with a release review that includes the code, its failures and the operator workload.

Each lesson explains one mechanism, shows it in a diagram, gives you a concrete exercise and then changes an assumption. Ten projects connect those mechanisms into a reviewable implementation.

One case, from evidence to accepted action

The whole course fits in one picture. Put a finger on any stage and you can say what the case looks like at that moment: which policy governs it, what the model has proposed, what a person approved and what the service has committed. Stage 6 is the one the course turns on.

Throughout, you keep a suggested replacement, an approved request and an accepted service receipt apart, and a receipt never means the item shipped. That distinction runs through the diagrams, fixtures, tests and capstone.

Start free, then build the complete system

Read three complete lessons without an account:

• Build a support agent around one permitted outcome: decide what the model may propose and what code must enforce.
• Define agent success before writing the prompt: turn a customer outcome into an acceptance check.
• Read an agent evaluation without losing the failures: inspect a worked release comparison where the better average hides an unsafe action.

Three more lessons unlock with a free account: build the baseline, select applicable policy evidence and design a scoped read tool. Their examples run independently.

Premium opens twenty-two further lessons: context validation, adversarial approval tests, crash recovery, evaluator calibration, paired release analysis, a resettable agent gym, error analysis, production judge monitoring and the capstone. The downloadable companion includes original fixtures, a reference implementation, learner exercises, deliberate test mutations and a release-review worksheet. It requires active paid Premium or approved complimentary download access. Referral Premium includes the online lessons.

Build the evidence before the training run

The evaluation dataset project turns failure notes into acceptance predicates, named coverage cells and source-family splits. Its Python checks detect missing critical cases and declared source families crossing splits. The free evaluation guide connects this work to judges, repeated trials, release statistics and live monitoring.

Post-training: from traces to a checkpoint review

Five Premium lessons connect the existing gym to model improvement. Audit assistant loss masks, distinguish teacher feedback from verified outcomes, calculate a group-relative reward signal and validate rollout provenance. Two projects supply runnable Python checks and a release packet to review. An optional CPU script executes two real optimizer steps on a tiny random model; it validates the pipeline without claiming trained-agent quality.

Start with what training should change, or read the free practical introduction. The core checks need no model API or GPU. The optional optimizer script needs a separate Python environment and downloaded training dependencies.

Frameworks you can reason about

Dated September 2026 field notes and a companion migration lab cover Google ADK 2.0, Anthropic's Claude Agent SDK, LangGraph, OpenAI Agents SDK and Pydantic AI. You will map workflow nodes, tools, approval pauses, retries and stored state onto the same support contract. The durable lessons remain useful when SDK syntax changes.

The supplied implementation runs offline with recorded model decisions and tests control behavior, including actual worker process death. The optional framework lab specifies the additional live-runtime checks to perform; it does not claim those integrations have already passed. All cases and evaluation outputs are synthetic, with no customer data or paid model account required for the core course.

Where this fits

Use FDE Foundations for the role and first customer engagement. Use The FDE Engagement for the complete delivery lifecycle. This course gives you a connected agent implementation to explain and inspect; Hard Deployments extends the operating constraints afterward.

On your first pass, work in order, and save your acceptance contract, failing tests and release decision as you go. The final artifact is a system you can defend with evidence, including an honest account of what remains untested.

Questions before you start

Is this AI agents course suitable for beginners?

It suits beginners to agent engineering who can already write Python functions, work with dictionaries and handle exceptions. You should recognize JSON and HTTP requests. The lessons introduce agent control, recovery and evaluation from the beginning, then build toward an intermediate capstone. If you are new to the forward deployed engineer role itself, begin with FDE Foundations.

What will I build in the ten projects?

You will build a reproducible support baseline, repair its evidence path, test its authority boundary, recover an interrupted write, compare two releases, build an agent gym, construct a versioned evaluation dataset, audit training traces with an optional real SFT smoke run, validate rollout admission and assemble the final support system. Each project supplies the records and expected behavior needed to begin. The capstone asks you to defend the implementation using tests and a record of what remains unverified.

Does the course teach Google ADK 2.0 and Anthropic's agent framework?

Yes. The September 2026 field notes and Premium companion map the course's contracts to Google ADK 2.0 and Anthropic's Claude Agent SDK, alongside LangGraph, OpenAI Agents SDK and Pydantic AI. The reference implementation uses Python's standard library. Framework migration exercises specify the live checks you must run before relying on a port; they are not pretested SDK integrations.

Do I need a paid model API or a GPU?

No paid model API, API key or GPU is needed for the supplied local projects. They run with Python 3.10 or later and recorded model decisions. Optional live framework experiments need the chosen runtime and any provider access it requires. The local checks establish control behavior; they do not measure a live model's answer quality.

Which lessons are free, and what does Premium include?

Three lessons are public, and a free account opens three more. Premium opens the remaining twenty-two lessons. The downloadable project companion requires active paid Premium or approved complimentary download access and includes recovery exercises, adversarial checks, evaluation fixtures and the release-review worksheet. Referral Premium covers the online lessons. See current access terms and pricing, or start the first free lesson.