Your entire test pyramid rests on a promise agents do not keep: same input, same output. So how do you ship something you cannot assert on, and how do you tell an auditor, a customer, or your own on-call engineer that it works?

This is the question every enterprise asks second, and almost no conference answers.

We will build a strategy in layers. Deterministic first: record model responses and test your orchestration like ordinary software, because most of your bugs are in the plumbing rather than the model. Then behavioural: table-driven cases over what the agent decided and which tools it reached for, asserting on the shape of the action rather than the wording of the prose. Then evaluation sets: how to build one from real traffic without creating a data protection incident, how large it has to be before the number means anything, and how to decide what score is good enough to ship.

Then the uncomfortable part. Your most important dependency is a model that can change underneath you without a version bump, so regression detection is not optional and cannot be manual.

Finally, production. Trace-level observability across multi-step runs, telling a bad answer apart from a broken tool, sampling for human review at a cost you can defend, and which metrics actually predict user complaints. Plus the honest limits: what you cannot test, and how to design a system whose worst case is acceptable rather than one whose average case is impressive.

Takeaways

- A layered strategy: deterministic orchestration tests, behavioural assertions, evaluation sets
- Building an evaluation set from production traffic without a data protection problem
- Detecting regressions when the model is an unversioned dependency
- Observability and alerting for multi-step agent runs
- Designing for an acceptable worst case instead of an impressive average

Preferred duration: 45 minutes including Q&A. Can be delivered in 30 or 60 minutes on request.

Target audience: senior developers, architects, QA and test engineers, and platform teams responsible for AI features in production. No machine learning background required.

Level: intermediate.

Language: the patterns are language-agnostic; code examples are in Go.

Technical requirements: my own laptop (USB-C / HDMI) and internet access for the live demo. A recorded fallback is always available.

First public delivery: not yet delivered.