AI engineering has stopped being about models and started being about systems. That is the central message from the QCon AI New York 2026 program, which has confirmed 23 sessions ahead of its December 15-16 run at The Westin Jersey City Newport. The conference targets senior engineers, architects, and technical leaders who are building AI systems that must survive contact with real infrastructure.
When Agents Stop Being Tools and Start Being Users
Nancy Wang, CTO at 1Password, will deliver the keynote titled When Software Becomes a User: Identity and Authorization for Agents in Production. The talk addresses a problem that has no precedent in traditional security engineering: software agents that act on behalf of multiple people, invoke tools nobody listed ahead of time, spawn subagents, and keep running without a human watching over them.
Conventional identity systems rest on a simple set of assumptions. There is one human principal. The role stays stable. The session has boundaries. And afterward, someone can be held responsible. Agents break every one of those assumptions. Wang's keynote will examine how to grant delegated authority across chains of agents, how to maintain auditability through multi-hop tool calls, and how to give an agent just enough access to finish a job without ever dropping credentials or secrets into its context. She will also address the two common failures: service accounts that carry too much privilege for too long, and permission sets so narrow that the agent simply cannot do its work.
Engineering Controls for Agents That Act on Infrastructure
LinkedIn operates a Kubernetes-based compute platform covering more than 500,000 nodes and roughly five million pods. Ronak Nathani, Principal Staff Software Engineer, leads a team that built an operations agent capable of investigating deployment failures through Slack, coding-agent plugins, and automated workflows.
The session, Inside LinkedIn's Kubernetes Ops Agent: Skills, Tools, and Guardrails, focuses on what happens when an agent moves from explaining a problem to acting on production infrastructure. The team has implemented server-side rate limits, protections around delete and scale-down operations, access controls, bounded actions, and peer approval for production changes. Nathani will also describe how recurring operational work gets converted into reusable skills, and how JIRA history and prior support conversations are used to identify problems those skills can handle.
One Platform for 300 Models at Netflix
Netflix has spent five years consolidating its model-serving infrastructure into a single multi-tenant platform. The result handles approximately one million inference requests per second across more than 300 models, serving everything from recommendations and commerce to newer large language model applications.
Staff Software Engineer Rajat Shah will present One Infrastructure, Every Model: How Netflix Scales ML & GenAI. The talk covers architectural and organizational trade-offs that come with forcing models with different operational needs onto one platform. Latency requirements range from tens of milliseconds to over 300 milliseconds. A common deployment contract had to be defined. Teams had to decide where business logic ends and model logic begins.
For ML practitioners, the change was practical: instead of rebuilding deployment and serving infrastructure for each model, teams work against a shared platform and focus on the modeling problem itself. Shah will also address the operational costs and ongoing challenges this consolidation introduced.
Measuring Whether an AI System Actually Works After Shipping
DoorDash's Bruna Pereira will present After It Works: Trusting and Teaching Alchemy, covering the company's content-agnostic AI moderation platform. Alchemy uses a lower-cost in-house classifier to gate access to a more expensive large language model. About 90% of content classified as clearly acceptable never reaches the costlier layers.
The talk focuses on the difficult work that follows the first production release. When there is no clean ground truth, how do you determine whether a decision was correct? How do you evaluate judgments that are inherently nondeterministic? How do you change prompts or models without losing visibility into what changed? Pereira will describe an evaluation harness supporting shadow-mode testing, backtesting against historical production data, labeling workflows, and metrics tied to incident reduction rather than model accuracy in isolation. She will also explain how LLM judgments feed back as training data for the cheaper classifier, reducing latency and inference costs over time.
What the Program Signals
Conference Chair Hien Luu framed the shift plainly: "AI engineering has become systems engineering." The challenge has moved from model behavior to system behavior. That means giving agents bounded execution authority, managing context and state, and wrapping probabilistic models in deterministic control planes. Harness engineering, continuous evaluation, observability, and policy enforcement are now core infrastructure. Inference economics, including latency, token usage, model routing, and cost, have become first-class architectural constraints.
Taken together, the confirmed sessions point to AI engineering drawing heavily on disciplines that have long existed in distributed systems, security, platform engineering, and site reliability. The questions are no longer speculative. Who or what is authorized to act? How are those actions constrained? Where should shared infrastructure be introduced? How do teams determine whether a deployed system continues to behave acceptably? QCon AI New York will spend two days answering them with concrete engineering examples from companies that have already built and deployed these systems.
Registration is open at the conference website. Early bird pricing ends October 13.