people · May 22, 2026

Arize AI Publishes Self-Improving Agent Demo Using Context Graph of Human Disagreements

Arize AI published a May 19, 2026 blog post detailing a procurement agent demo with 130 purchase requests. A simulated reviewer named Vera Fye overrode the agent in 60 cases for a 53.8% baseline match rate. After four cycles of mining overrides into a context graph and updating runtime config via Claude Agent SDK, the agent reached 83.1% match with Vera on the same 130 requests without source code changes or fine-tuning. The demo uses LangChain for the agent, traces in Arize AX, and patterns like DataStream cost-overrun factors and CloudBase CTO relationships extracted from 130 sessions.