Day 2 Lightning Talks
summary
A 5‑minute lightning talk that describes practical observability for agentic AI: a financial agent system is instrumented to surface failures across three-layer observability (task, reasoning, execution), with metrics like retries, token usage, and tool invocations logged. Prompts are privacy‑protected via hashed prompt envelopes + event replay so engineers can reproduce and debug decision paths without exposing PII. Mentions integration with trackers such as Langsmith and PromptFlow and moving from opaque agents to debuggable pipelines.
AI-written summary · about
