Lessons Learned Building Agentic Systems

Models keep getting better, but the systems we build on them keep getting harder to make reliable. Applications now carry far more context, hand work between multiple agents, and act with more autonomy against real systems and real customer data. Each of those adds new ways to fail.

In this session, four senior technical leaders from LendingTree, Comcast, ReFocus AI and Civio share the lessons they learned the hard way building agentic systems that real users depend on.

What you’ll learn

• Practical patterns for assembling the right context for the right task, person and moment, and why context engineering trumps prompt engineering.
• How to keep state and conversation context intact across agent handoffs, and why guardrails need that context too (without adding latency).
• How to design for failure instead of the happy path: recovery, validation, and treating autonomy as a bounded capability, with deterministic code and humans where correctness matters most.
• How to move evaluation and observability into the main loop: trace logging, curated eval datasets, LLM-as-judge scoring, and measuring the system rather than the model.
• How to give agents controlled access to real systems and enterprise data: authority, memory, governance and least privilege.
• Why the model is becoming the least differentiated part of the system, and how shared capabilities and infrastructure make the fifth agent faster to ship than the first.

Speakers

• Eric Hanson, AI Platform Architect, LendingTree
• Bhavin Parikh, Senior Principal AI Architect, Comcast
• Abbas Raza, CTO, ReFocus AI
• Anindo Banerjea, CTO, Civio
• Tony Karrer (moderator), founder of TechEmpower and cofounder of the LA CTO Forum

Members: the Zoom link is in your calendar invite.