Frontier AI Summit Europe 2026 — Schedule

Serving a model from Frankfurt to a user in Lisbon spends milliseconds before the first token exists, and most reference architectures assume that cost away. This talk walks through moving a production inference path onto EU-only capacity: where the region boundaries actually fall, what data residency does to a caching layer designed without it, and which optimizations paid for their operational cost. We cover routing across three regions, speculative decoding on smaller hardware, and the degradation path that keeps a page usable while one region is unhealthy. Every number comes off a live service, including the migration week where p99 got worse before it got better and the rollback plan we spent two days writing and never used.

Most teams inherit at least one agent they did not write: a vendor product, a contractor's prototype, or a colleague's weekend build that is now load-bearing. This workshop is about getting a defensible read on one of those without access to the prompts or the training data. You will start from a black-box agent, build a trace harness around it, and design an eval set from its observed failures rather than from its documentation. We cover sampling real traffic without leaking it, writing assertions that can actually fail, and estimating cost per resolved task. Bring a laptop with Python 3.11 and Docker. You will leave with a running harness, a scored baseline for the agent you brought, and a short list of the failures its vendor page does not mention.