Agentic AI
Major tech firms reveal agent, eval, and inference strategies at QCon AI New York
Engineers from LinkedIn, Netflix, and DoorDash share production lessons on securing agents, running evaluations, and scaling inference workloads.
InfoQ
QCon AI New York 2026 has unveiled a comprehensive program for December 15–16 featuring hands-on sessions from practitioners at leading companies building production AI systems. The conference explores the operational and architectural choices that teams face when moving beyond initial deployments—particularly around agent authorization, evaluation frameworks, model serving, and AI-assisted development workflows.
Securing agents in production#
One core challenge runs through multiple sessions: how to grant agents the permissions they need while maintaining safety and auditability. Nancy Wang, CTO at 1Password, opens the conference arguing that traditional identity systems break down for software agents, which lack the stable roles, bounded sessions, and human accountability that credential schemes assume.
Practical implementations follow across the program. Figma's sandbox and caching team discovered that trust boundaries for agents extend well beyond initial isolation layers. By using warm pools to pre-provision resources, they cut startup latency by 90% and lifted availability from 99% to 99.95%. LinkedIn's approach to its Kubernetes operations agent—which coordinates activity across 500,000+ nodes and millions of pods—places guardrails server-side, enforcing rate limits, preventing destructive operations, and requiring peer sign-off for production changes.
Making evaluations predict real-world impact#
Companies are connecting offline testing to production behavior through calibrated evaluation loops. DoorDash's agentic ordering system uses an LLM judge trained against human raters to score every proposed change before launch. By applying the same rubrics to both simulated and actual sessions, teams can trust that improvements in testing translate to customer-facing gains.
Similar challenges appear across organizations. Google DeepMind's infrastructure team highlights a recurring problem: evaluations that pass initially, then fail on subsequent runs. The talk addresses building evals that remain stable and accessible enough for engineering teams to maintain.
Handling inference at scale#
Model serving emerges as a critical infrastructure challenge. Netflix's unified platform processes roughly 1 million inference requests per second across over 300 models after five years of consolidation work. The sessions cover the operational trade-offs of multi-tenancy and how to evolve serving platforms over time.

