The program for QCon AI New York (Dec 15‑16 2026) now lists sessions where engineers from LinkedIn, Netflix, DoorDash, Figma, Pinterest, Honeycomb, Harvey, Google DeepMind and OpenAI discuss the concrete ways they operate AI agents, evaluation pipelines, and inference workloads in live environments. Practitioners should note that these talks move beyond research demos to production‑grade concerns that directly affect platform design, cost management, and security posture.
Agent Permissions and Guardrails
Presentations emphasize the need for explicit permission models and runtime guardrails around autonomous agents. For teams building or integrating AI agents, this suggests a shift toward defining scoped capabilities, monitoring actions, and enforcing policy checks as part of the deployment pipeline.
Evaluation Loops and Inference Cost
Speakers highlight continuous evaluation loops that feed back model performance and the importance of tracking inference expenses. Practitioners should consider instrumenting metrics for latency, throughput, and cost, and designing feedback mechanisms that can trigger model retraining or scaling decisions without manual intervention.
AI‑Driven Code Review and Team Adoption
Sessions on AI‑assisted code review illustrate how organizations are embedding model‑based analysis into their CI/CD workflows. This points to a broader trend of automating quality gates and requires teams to evaluate integration points, false‑positive handling, and the impact on developer velocity.
Related CloudNinjas coverage: AI engineering.
What This Means For Practitioners
Engineers should start inventorying current AI workloads for permission boundaries, cost visibility, and automated review steps. Evaluating existing tooling against the patterns discussed can surface gaps in observability, policy enforcement, and scaling strategy. Ongoing monitoring of QCon follow‑up sessions will provide deeper guidance as these practices mature.
