AI agents are changing more than the way code is written. They are becoming users of production systems, initiating customer actions, querying observability data, and influencing how software is tested and released. That creates practical questions for senior engineers. Which decisions can safely be delegated to an agent? What evidence is needed before agent-generated work reaches production? How should existing systems expose context without expanding risk? Where must experienced engineers retain direct control?
QCon San Francisco 2026, taking place November 16–20, brings together practitioners from Airbnb, OpenAI, Netflix, Honeycomb, and other engineering organizations to share how they are answering these questions in production. The program connects emerging AI practices with established lessons from distributed systems, architecture, data platforms, observability, and resilient operations.
Moving AI agents into customer-facing systems
An agent that drafts text presents a different level of risk from one that can take action on a customer’s account. Teams building these systems need controls that go beyond a prompt or a single model-level safeguard.
In "How Airbnb Guardrailed Its AI Customer Support Agent", Weiping Peng, Distinguished Engineer at Airbnb, will examine the safeguards behind an agent that serves millions of customers, maintains context across conversations, and can initiate account actions.
Weiping will discuss the layered approach used to prepare the agent for production, including input sanitization, classifiers, shadow testing, false-positive management, and rapid-response mitigations. For engineers introducing agents into customer-facing workflows, the session offers a concrete example of how preventive controls, offline testing, and production response can work together.
Deciding what coding agents should and should not own
Coding agents can reduce the time required to implement and iterate on software, but speed does not remove the need for engineering judgment. Teams still need to decide how generated work is verified and who owns architecture, quality, and release decisions.
Brian Yang, Member of Technical Staff at OpenAI, will address these boundaries in "Lessons from Building a $100M Product in Six Weeks at OpenAI". Brian will share the operating model used to build OpenAI Ads with coding agents from the beginning and scale it to more than $100 million in annual recurring revenue in under six weeks.
The session will cover feedback loops, verification, token economics, and the decisions that remained under human ownership. For senior engineers and technical leaders, the case study provides a way to evaluate where coding agents can accelerate delivery and where experienced judgment remains difficult to delegate.
Understanding system-wide AI trade-offs at scale
While AI is changing development workflows, teams still face familiar distributed-systems problems: latency, capacity, compatibility, failure handling, and safe rollout. Optimizing one part of a system can create costs elsewhere, especially at high request volumes.
In "Orderly Keys, Wild Values: Adaptive Compression for Distributed Key-Value Storage", Netflix engineers Joseph Lynch and Ayushi Singh will examine compression as a systems problem across billions of daily requests and petabytes of key-value data.
Their session will explore the trade-offs among storage footprint, cache behavior, network I/O, p99 latency, dictionary versioning, compatibility, and rollout safety. Engineers responsible for high-throughput data systems will see how a seemingly local optimization can affect the behavior and risk profile of an entire platform.
The Distributed Systems in Production track will extend this discussion across latency, consistency, observability, capacity, and failure handling. Khawaja Shams, QCon San Francisco 2026 Program Committee member and co-founder and CEO of Momento, said:
"Vidhya Arvind, Track Host and Tech Lead & a Founding Architect for the Data Abstraction Platform @Netflix, has put together an incredible program covering the parts of distributed systems that matter most in production. You will learn about taming latency, consistency, failure handling, observability, capacity, and the operational tradeoffs required to run at scale."
Making production systems understandable to agents
As agents begin querying and operating production systems, teams need to think about how those systems communicate intent, context, and constraints to non-human users. An API may be technically accessible to an agent without being understandable enough to produce reliable results.
In "Making Production Legible to Agents: Lessons From Building an Observability MCP", Honeycomb Technical Fellow Liz Fong-Jones will share lessons from building an MCP server used by more than 40% of Honeycomb’s weekly active users to run production queries through agents.
The session will cover token economy, tool descriptions, schemas, evaluations, output formats, and defects found through real-world use. The lessons apply beyond observability: they show how teams can make existing production systems more useful to agents while reducing ambiguity in the actions those agents take.
Applying the conference to current architecture decisions
Attendees who want to work through the implications for their own systems can join the InfoQ Certified Architect Program. The four-day experience combines the three conference days with a dedicated peer cohort and a half-day workshop on November 19.
The cohort will meet before and during QCon to compare how the ideas presented apply across different organizations. In the final workshop, led by O’Reilly author and former AWS and DAZN architect Luca Mezzalira, participants will bring a current architecture challenge and examine it with the group. Participants earn the ICSAET certification after completing the program.
Optional hands-on training continues on November 19–20. Workshops cover choosing an SDK for AI agents, AI-native development with harness engineering, working with Claude Code and Codex, measuring developer productivity, incident response and learning, and AI-assisted coding workflows. Training can be added to a conference pass or attended separately.
For senior engineers, architects, and engineering leaders, the QCon San Francisco program provides an opportunity to compare emerging AI practices with the production lessons teams have learned from distributed systems, platform engineering, observability, and operations. The focus is not on what agents may eventually do, but on the engineering decisions teams are already making as these systems reach production.
QCon San Francisco 2026 takes place at the Hyatt Regency San Francisco. The conference runs November 16–18, followed by training on November 19–20. Early-bird conference tickets are available for $2,955 through October 13.
Explore the production engineering sessions at QCon San Francisco 2026.