Why AI Agents Are Blowing Up Your Telemetry Bill with Nikhil Mungel

<p>Most conversations about AI agents stay on what agents can do. Less attention goes to what they leave behind: logs, traces, and tool-call chains at a volume nobody's pipeline was built for. Nikhil Mungel, Head of AI R&amp;D at Cribl, talks with Kate Holterhoff about why that is becoming observability's cost and noise problem. They discuss why token prices keep falling while total spend climbs, how employee spend differs from the tokens a product burns serving customers, credit-based versus outcome-based billing, why users always reach for the priciest model, and how the Open Cybersecurity Schema Framework (OCSF) standardizes what agents emit. Underneath all of it sits the question the industry keeps dodging, which is whether to keep scaling the backend or get stricter about what gets collected at all.<br /><br />This RedMonk conversation is sponsored by Cribl.<br /><br />Show notes: <a rel="noopener noreferrer nofollow" href="https://redmonk.com/videos/nikhil-mungel/" target="_blank">https://redmonk.com/videos/nikhil-mungel/</a><br /><br />Chapters<br />00:25 Leading AI research and development at Cribl<br />01:46 Token prices fell, token bills grew<br />03:39 The research agent that thought all night<br />06:54 More software, more agents, more telemetry<br />09:37 Employee spend versus cost of goods sold<br />12:05 Credits, outcomes, and other ways to bill for agents<br />14:46 Shadow AI and the cloud governance rerun<br />17:00 Writing policy: cost, access, and blast radius<br />17:49 Local models and regulated industries<br />18:40 What Cribl collects, plus Cribl Guard<br />21:24 Token gateways as the enforcement point<br />24:56 Model anxiety and automatic routing<br />27:29 What OCSF is and why it caught on<br />30:15 Observability as an information compression problem<br />33:48 Zero data retention versus model safety<br />35:45 Should agents know what they cost?<br />36:35 Where to find Nikhil and Cribl's research</p>