QCon AI New York 2026 brings together production patterns for agents, evals, and LLM cost

When the hype becomes operations

Once an AI agent moves out of the prototype stage and into production, the problem changes shape. Generated code arrives faster than review can keep up with, token costs rise, evals need to be trustworthy enough to decide a deploy, and the agent needs permissions and guardrails like any other system that touches production. QCon AI New York 2026 has confirmed sessions from senior engineers who have already faced these decisions in systems they themselves keep running, according to InfoQ.

The event takes place on December 15 and 16, 2026, at The Westin Jersey City Newport, about a five-minute PATH train ride from the World Trade Center, and is aimed at senior engineers and technical leaders who make decisions about AI in production. The idea, according to the coverage, is not the generic keynote speech: it's the account of someone who has already made the call and has the numbers behind it for another team to compare against their own reality.

Identity and authorization: giving an agent production access

The opening keynote, by Nancy Wang (founder of the AWS Data Protection area), starts from a concrete dilemma: give the agent a long-lived service account and hope nothing goes wrong, or restrict access so much that it can't finish the task. The talk When Software Becomes a User: Identity and Authorization for Agents in Production proposes alternatives to these two extremes.

Anyone who wants to see this model working on a real platform has the session by Ronak Nathani, who led LinkedIn's migration to Kubernetes. In Inside LinkedIn's Kubernetes Ops Agent: Skills, Tools, and Guardrails, he details an operations agent that acts across more than 500,000 nodes and 5 million pods, with server-side guardrails: call rate limits and approval from a human peer before any change to production. Together, the two talks deliver an authorization model and an already-tested set of guardrails to compare against what the team has today.

Evals that decide whether an agent goes live, and how much it costs to run it

An eval only really works when the team trusts the score enough to decide whether to ship or hold back a change based on it. Fabio Flores, who leads foundational development of Agentic Ordering at DoorDash, will argue along these lines that the same judge should evaluate simulated sessions and production sessions, so that an improvement measured before launch still holds after it. The talk is called From Signal To Impact: The Eval Loop Behind DoorDash's Agentic Ordering System.

The other side of the equation is cost. Agent workloads that worked fine as a prototype become too expensive in production as context grows. Tejas Chopra, who worked on distributed cache infrastructure at Netflix, will show in Make Each Token Count: Compression, Caching, and Retrieval for Production LLM Context how prompt compression and ordering increase the provider's cache hits, along with tiered storage for conversation history and the failure modes his team encountered along the way.

When AI writes code faster than the team can review it

Human review becomes the bottleneck as AI tools generate more code. Min Chen, who leads the adoption of AI-native software engineering at LinkedIn, will argue that the team should filter out low-confidence findings before they reach the developer. Her session, High-Signal AI Code Review That Adapts to Your Codebase at Scale, describes a system that reviews tens of thousands of pull requests per week across about 10,000 repositories, and how the team measures whether suggestions actually make it into the code that ships to production.

More code, faster, doesn't help every organization in the same way. Nathen Harvey, who leads DORA research at Google Cloud, proposes in Building in the AI-First World: Scale Brilliance, Not Bottlenecks measuring AI maturity beyond adoption metrics, identifying which pre-existing problem AI tends to make worse. His central thesis, according to InfoQ, is that AI amplifies both what the organization already does well and what it already does poorly.

What changes for those who build this day to day

For an engineering team that already runs, or is about to run, agents in production, including in Brazil, these five sessions cover decisions that are normally made on the fly: what identity model the agent uses instead of a generic service account, whether the eval that validates a prototype is the same one that validates production, how token cost enters as an engineering variable from the design stage, and how to triage the volume of AI suggestions before it becomes noise for whoever reviews it.

The QCon AI Program Committee invites senior professionals with something other teams can learn from, and the talks are vendor-neutral, with no hidden product pitch: each one is an account of what a team built and why. With a ratio of one speaker for every eleven attendees, there's still room for a follow-up question to whoever actually built the system. All confirmed sessions are listed on the QCon AI New York 2026 program page.