Everyone can build an agent. Shipping one you can trust is the hard part - agents never behave the same way twice, so "looks good on my machine" doesn't survive production.

This hands-on workshop is for engineers past the demo phase. We treat a multi-agent system like real software: something you evaluate, gate, and ship on a pipeline - not something you eyeball and hope.

We build a multi-agent system, then put it on rails:

- Evaluation for non-deterministic output: how to measure "good" when the answer changes every run, and build eval sets you actually trust.
- Gate-keeping: the checks an agent must pass before it's allowed to act or ship, and where human-in-the-loop belongs.
- Frameworks and tooling: what to actually reach for - orchestration, eval harnesses, and observability.
- CI/CD for agents: wiring evals and gates into a pipeline so regressions get caught automatically, not in prod.
- Red teaming as a gate: prompt-injection and goal-hijack checks that run inside the pipeline, one gate among many.

You leave with a working system that evaluates itself, blocks its own bad releases, and ships on a pipeline you understand.