AI Systems as Critical Infrastructure: QCon AI Unveils 2026 Program on Agent Authorization, Production Safety

QCon AI New York is shaping its 2026 lineup around a fundamental shift in how engineers approach AI: moving away from optimizing isolated models toward building robust, controlled systems that can operate reliably in production at massive scale. The conference, running December 15–16 at The Westin Jersey City Newport, has now confirmed 23 of more than 30 sessions that reflect this evolution.

From Model Optimization to Systems Control#

Conference Chair Hien Luu describes the transition plainly: "AI engineering has become systems engineering." Rather than tuning model behavior in isolation, teams now grapple with how autonomous agents behave within broader architectures—granting them bounded execution authority, managing their state and context, and wrapping probabilistic models in deterministic controls.

This reframing elevates concerns that belonged traditionally to distributed systems and security: observability, policy enforcement, cost management, and auditability all become central architectural responsibilities. Latency, token usage, and inference economics shift from afterthoughts to primary design constraints.

Identity and Access Control for Software Agents#

Nancy Wang, CTO of 1Password, will keynote on the identity crisis that emerges as software agents become active participants in production systems. Traditional authorization assumes a human operator with a stable role and bounded session; agents instead operate across multiple users, call unenumerated tools, spawn sub-agents, and run unsupervised.

Wang's talk explores how to establish delegated authority chains, audit multi-hop tool invocations, and grant agents sufficient permissions without embedding secrets directly in their context—walking the line between unrestricted service accounts and constraints so tight that agents cannot accomplish their work.

Guardrails for Large-Scale Operations#

LinkedIn's Ronak Nathani will share lessons from deploying an operations agent against the company's Kubernetes infrastructure spanning over 500,000 nodes. The agent, integrated into Slack and developer tools, diagnoses deployment failures and can take corrective actions—but only behind carefully designed guardrails.

The presentation covers server-side rate limits, protections around destructive operations, access controls, bounded actions, and peer-review workflows for production changes. Nathani will also detail how the team converts recurring operational patterns into reusable agent skills by mining JIRA history and support logs.