QCon San Francisco 2026 tackles guardrails, boundaries, and system design for AI agents in production

As AI agents move beyond experimentation into customer-facing production systems, they raise urgent design questions that no prompt engineering can fully solve. How much autonomy can engineers safely grant an agent that handles customer accounts? What tests separate reliable automation from costly mistakes? QCon San Francisco 2026 (November 16–20) assembles senior engineers from Airbnb, OpenAI, Netflix, Honeycomb and others to share concrete answers based on systems already running at scale.

Guarding agents in customer-facing roles#

An agent that generates text carries far less operational risk than one authorized to modify a user's account or initiate transactions. Airbnb's Distinguished Engineer Weiping Peng will walk through the multi-layer safeguard strategy behind the company's AI customer support agent, which now serves millions of users while maintaining conversation history and performing account actions. His talk covers input validation, classifier checks, shadow testing (running the agent offline to verify behavior before it reaches real customers), false-positive filtering, and rapid intervention procedures. The session is designed for teams deciding how to move their first agents into live customer workflows—a case study in building controls that don't slow down helpful automation but catch dangerous edge cases.

Boundaries for code-generation agents#

Speed matters, but so does judgment. Brian Yang, a technical leader at OpenAI, will describe how his team built and shipped the OpenAI Ads product with coding agents embedded from day one, reaching over $100 million in annual revenue within six weeks. The catch: humans remained decision-makers on architecture, quality standards, and release timing. Yang's talk covers feedback loops, code verification, token budgeting, and—crucially—which choices stayed under human ownership. For technical leaders, it's a blueprint for knowing when to let agents iterate and when to insert human review.

Production systems thinking at massive scale#

Distributing AI computation across global infrastructure surfaces the same old problems in new forms: latency spikes, cache efficiency, network overhead, and safe rollout. Netflix engineers will present a case study in compression as a systems problem across billions of daily queries and petabytes of data, examining how optimizing storage footprint can ripple through cache behavior and p99 response times. The broader Distributed Systems in Production track digs into taming latency, consistency, observability, and failure handling as AI workloads scale.