From Agent Authorization to AI Production Evaluation: QCon AI New York 2026

QCon AI New York has confirmed 23 of more than 30 planned sessions for its conference on December 15-16, 2026, at The Westin Jersey City Newport. The newly published talks cover identity and authorization for autonomous agents, guardrails for an operations agent running against large-scale Kubernetes infrastructure, shared model-serving platforms, and evaluation of AI decision systems after deployment.

Conference chair Hien Luu said the program reflects a shift in AI engineering work: AI engineering has become systems engineering, with the challenge moving from model behavior to system behavior. He said this includes giving agents bounded execution authority, managing context and state, and wrapping probabilistic models in deterministic control planes, and that harness engineering, continuous evaluation, observability, and policy enforcement are becoming core infrastructure. He added that inference economics such as latency, token usage, model routing, and cost are now architectural constraints.

In the keynote When Software Becomes a User: Identity and Authorization for Agents in Production, Nancy Wang, CTO at 1Password, will examine how identity and authorization models change when software agents become active users of production systems, covering delegated authority across chains of agents, auditability across multi-hop tool calls, and giving agents enough access without placing credentials or secrets in their context.

Ronak Nathani, principal staff software engineer at LinkedIn, will present Inside LinkedIn's Kubernetes Ops Agent: Skills, Tools, and Guardrails. LinkedIn's Kubernetes-based compute platform spans more than 500,000 nodes and five million pods, and its operations agent can be used through Slack, coding-agent plugins, and automated workflows for investigating deployment failures. The session covers controls including server-side rate limits, protection around delete and scale-down operations, access controls, bounded actions, and peer approval for production changes.

Netflix staff software engineer Rajat Shah will discuss One Infrastructure, Every Model: How Netflix Scales ML & GenAI, covering a five-year consolidation effort. The platform handles approximately one million inference requests per second across more than 300 models, supports latency regimes from tens of milliseconds to more than 300 milliseconds, and is used for recommendations, commerce, and newer large language model applications.

Bruna Pereira, software engineer at DoorDash, will present After It Works: Trusting and Teaching Alchemy, DoorDash's AI Moderation Platform. Approximately 90% of content classified as clearly acceptable never reaches the more expensive layers. The session covers evaluating nondeterministic judgments, an evaluation harness supporting shadow-mode testing and backtesting against historical production data, and using LLM judgments as training data for the lower-cost classifier.

Current early bird pricing ends October 13. The complete program published so far and registration details are available on the conference website.