When Your "New Teammate" Hallucinates: Observability as the Safety Net for AI Code
summary
Argues that observability must be baked into AI-generated code by default and shows a working feedback loop: an agent uses repo skills and a pipeline skill to plan, deploy (AWS Lambda), run a load test (336 requests in 90s), pull traces/metrics via DDCTL/API, diagnose issues (5s latencies, 502s, missing pagination), patch code (adds semaphores, spans, structured logging), redeploy and iterate — one session ran for almost 12 hours and continually improved behavior.
AI-written summary · about
chapters (11)
- 0:00AI code safety intro
- 1:59AI code needs guardrails
- 3:44Layered AI coding agents
- 6:34Planning before coding
- 9:01Skills enforce standards
- 11:11Observability for agents
- 13:18Runtime truth vs reviews
- 14:43AI feedback loop
- 16:29Live demo feedback pipeline
- 22:30Key takeaways
- 23:54Vibe coding to AI engineering
related talks
28mWhere are we going with observability?
30mWhen AI Takes Over: DevOps Nightmare Unleashed
30mAI Agents for Incident Management
16mAI Agents for Incident Investigation: The Good, The Bad, and The Ugly
28mDO WE STILL NEED TO "OBSERVE" THE FUTURE OF AI & OBSERVABILITY
29mBuilding Trust in AI Agents
10mLightning Talk: Building AI That Ops Teams Actually Trust
21mConvention Over Hallucination: Harness Engineering for AI-Powered...