Once an AI system can retrieve internal documents, generate production code, or take independent actions through tools, the engineering challenges and associated risks increase. Teams need to know what the system accessed, which actions it was permitted to take, how its output was verified, and who remains accountable when it fails.
Those questions become harder in production. Coding agents move beyond an individual developer's machine and into CI. RAG pipelines encounter sensitive or outdated data. Agents operate across systems with different permissions and failure modes. A workflow that performed well in a controlled test can behave differently when it meets real users, operational constraints, and incidents.
On Wednesday, October 14, InfoQ will host "AI in Production: What Breaks, What Works, and Who Approves It?", a free 60-minute live panel with five practitioners who build, secure, and operate AI systems in production. Everyone who registers will receive the recording, so engineers who can't join live can watch it later.
Building production AI puts two pressures in tension. Platform and coding-agent teams want agents to move faster and do more. Privacy and security engineers want tighter limits on what those agents can see and touch. The discussion will explore where teams draw those lines, and the consequences of granting too much autonomy or imposing the wrong controls.
The panel will examine how practitioners approach:
- Agent autonomy: which actions agents can take independently and which still require human approval
- Verification: how teams check AI-generated changes before they enter CI or reach production
- Data exposure: which sensitive or regulated data an AI system can access and which threats need to be modeled first
- Production RAG and platforms: the architectural and operational practices that continue to work beyond a controlled demonstration
- Making the internal case: how engineers can explain the need for stronger controls before an incident forces the issue
The panel
Hien Luu chairs QCon AI New York 2026, has led machine-learning platform teams at DoorDash and Zoox, and is the author of MLOps with Ray. As the facilitator of the InfoQ Certified AI Engineering Program, he works with senior practitioners on the architecture behind production AI systems. He will bring a perspective spanning RAG pipelines, agents, AI platforms, and reliability.
Katharine Jarmul is a privacy and security specialist in machine learning and AI systems and the author of Practical Data Privacy (O'Reilly). As the facilitator of the InfoQ Certified AI Security & Privacy Engineering Program, she works with practitioners on sensitive-data handling, threat modeling, guardrails, observability, and governance. She will examine how teams control what AI systems can access and how they assess the risks.
Zichuan Xiong is a Principal at Thoughtworks whose work spans domain-driven design, data mesh, organizational change, and agentic systems. He co-facilitates the InfoQ Certified AI-Assisted Engineering Program and will examine the engineering harness around coding agents, including how teams verify agent-generated changes before they move through delivery workflows.
Premanand Chandrasekaran is a Market Tech Director at Thoughtworks with close to 30 years of experience leading software teams, and co-author of Domain-Driven Design with Java. He co-facilitates the InfoQ Certified AI-Assisted Engineering Program and will discuss the permissions, review gates, and governance decisions teams face when agents act across the software delivery lifecycle.
Lan Chu is an AI Tech Lead and Senior Data Scientist with more than seven years of experience building production data and machine-learning pipelines and more than three years building generative AI products. At QCon London 2026, she presented Beyond the Demo: RAG Is Easy, Production RAG Is Not, drawing on production AI systems built on more than 10,000 documents. She will bring a hands-on perspective on making retrieval reliable beyond a demonstration.
A 60-minute panel with live questions
The moderator and panelists will seed the 60-minute discussion by preparing the initial questions together, while attendees can shape the conversation by submitting their own.
Questions can be submitted in advance through the free-text field on the registration form or during the session using Zoom's Q&A. Rather than offering a single blueprint, the panel will surface the reasoning and tradeoffs behind different approaches so attendees can compare them with their own systems, risks, and constraints.
Registration is free for the "AI in Production - What Breaks, What Works, and Who Approves It?" webinar.
Event details
- Date: Wednesday, October 14, 2026
- Time: 4:00 PM London, 11:00 AM ET, 8:00 AM PT
- Duration: 60-minute panel discussion
- Format: Live online webinar, with attendee questions via the registration form & Zoom Q&A
- Recording: Shared with everyone who registers
- Cost: Free