Descriptions:
Jason Lopatecki, co-founder of Arize AI, delivers a technical deep-dive at the AI Engineer conference on how to build agents that improve themselves — moving observability from a human-facing dashboard activity to the fuel for autonomous self-correction loops. The central argument: traditional observability (logs, traces, graphs) was designed for humans to read; the next generation needs to be designed for agents to act on.
Lopatecki walks through Arize’s own journey building Alex, their in-product AI assistant, explaining how the team inverted the classic debug loop — instead of a human spotting an error and directing a fix, Signal (Arize’s new background agent) monitors production traces continuously, surfaces pattern-level issues before any human looks at them, and generates a proposed GitHub PR with a fix already attached. He illustrates this with a real production example involving a ‘stream canceled’ event that caused Alex to call to-do-update repeatedly in a loop — a failure class invisible to any static eval rubric.
The talk argues that the bottleneck in agent improvement today is not the fix itself but the cold-start cost: gathering enough evidence to confidently diagnose what went wrong. Signal addresses this by pre-loading context from traces and evals before a human engineer ever engages. Lopatecki also previews a broader vision where teams log and trace an order of magnitude more than today — because agents, unlike humans, can process that volume — enabling truly continuous self-improvement cycles.
📺 Source: AI Engineer · Published July 24, 2026
🏷️ Format: Deep Dive
