AI Engineer Summit 2026

A look at what changed in the last twelve months — from demos to systems, from prompts to products — and what the discipline needs to figure out next.

Agents fail in ways that traditional services do not: partially, plausibly, and expensively. This talk is the playbook we wish we had — bounded tool surfaces, idempotent actions, replayable traces, checkpointing, and the escalation paths that keep a human in the loop without keeping them in the way.

A working tour of a real evaluation harness: dataset curation, judge prompts you can defend, statistical significance on small samples, and wiring the whole thing into CI so a regression blocks a deploy rather than surprising a customer.

Multi-tenant inference from the operator's seat: scheduling, memory packing, cold-start mitigation, noisy-neighbour isolation, and the observability you need before you can safely oversubscribe anything.

Six months of session recordings, support tickets and telemetry from a shipped AI feature, and what they revealed: the prompts people really write, where they give up, the workarounds they invent, and the three changes that moved retention.

Four practitioners compare notes on the agent architectures they abandoned, the assumptions that did not survive production, and what they would build differently today.