23. Agentic Engineering - Observability

These slides are based on Chapter 23 of the book Agentic Engineering from First Principles, available at https://agentic-engineering-book.com/... Watch the full course in order:    • Agentic Engineering - Full Course   The question: How do we reconstruct what an agent actually did, and attribute quality, latency, cost and failure to the steps that caused them? By the end of this lecture, you will have built a nested trace. And you will know how to attribute cost and latency without double-counting, and without optimizing the wrong step. Chapter principle: Observe causality, not just activity. A useful agent trace explains which decision caused which work, what that work consumed, how state changed, and where quality or failure emerged. Chapters: 0:00 Title 0:22 The question for this lecture 0:45 Roadmap 1:17 A final answer hides the execution 1:44 Logging is not enough 2:18 Three signals, and the span 2:56 Experiment 41: the kernel's tracer 3:34 Experiment 41: a tiny traced agent 4:05 Experiment 41: what happened 4:41 Trace IDs and span IDs 5:11 Trace decisions, not hidden chain-of-thought 5:38 What belongs on a span? 6:21 Deliberately remove parent-child links 6:48 Trace delegation and state changes 7:22 Trace retries and fallbacks explicitly 7:45 Experiment 42: cost and latency 8:24 Experiment 42: what happened 9:00 Beware double-counting nested cost 9:29 The critical path 10:09 Which latency matters? 10:51 Deliberately optimize the wrong span 11:12 Metrics, sampling and redaction 12:02 A minimal trace schema 12:42 Chapter principle 12:58 Key takeaways 13:30 Try it yourself 14:15 Run the code 14:41 Next lecture Download the book: https://agentic-engineering-book.com/... Code for this lecture: https://github.com/agentic-engineerin... All the code: https://github.com/agentic-engineerin... Author, Christiaan van der Walt, PhD:   / christiaan-van-der-walt-phd-749b02a   Narrated with a synthetic voice.