JANUARY 12TH, 2027
12:00PM CST
VIRTUAL WEBINAR
About This Webinar
Observability has spent a decade getting very good at one question: did the request succeed? Agents changed the question. An agent can return a clean 200 in four seconds, call the wrong tool, read stale context, and produce a confidently incorrect answer. Nothing in a conventional stack registers that as a failure.
Two surveys make the gap concrete. In Splunk's State of Observability 2026, 56% of respondents said they cannot measure AI workload metrics such as model response quality. In a Cloud Security Alliance study with Aembit, only 28% of organizations could trace an agent's actions back to a human sponsor across all their environments.
Agent workloads need five signals that conventional instrumentation does not produce: who invoked them and on whose authority, what they cost, what they touched, what context they saw, and how good the answer was. In most organizations today, nobody owns any of the five. This session argues that agent observability fails at ownership first, which makes it a platform problem.
We use AWS as the worked example, including a live demonstration of a trace crossing into an agent call, going dark, and being lit back up. The platform of the future lets anyone say, for any agent in production, who asked, what it cost, what it touched, what it saw and whether it was right.
Webinar Itinerary
Green and Wrong: The Failure Your Stack Reports as Success: A composite agent returns clean status codes, normal latency and green dashboards while giving wrong answers for weeks.
The Five Signals: What Agents Need Your Platform to Emit: Caller identity, cost attribution, tool-call spans, context lineage and quality scores, and why the standards have not caught up.
Where the Trace Goes Dark: Lighting It Back Up on AWS: A live demonstration of a trace that ends at an agent call, then completes with OpenTelemetry and the AWS observability stack.
Who Owns the Score: Quality as a Reliability Signal: Continuous evaluation of sampled production traffic, with thresholds, alarms and error budgets.
Making Instrumentation the Default: Splitting the Work: How platform, product and security teams divide responsibilities, and what a golden path for agent workloads should emit.
Five Questions for Your Own Platform: A maturity check you can apply to any agent running in production today.

Who This Webinar is For
Platform and SRE Leaders
Own the observability stack and the on-call rotation for agent workloads. They need to know which signals every agent emits before it reaches production, and they want answers to who, what and why without another dashboard to maintain.
Product and Engineering Teams
Build agents into production services. They need a golden path that emits caller identity, cost, tool calls, context and quality scores before anyone writes agent code.
Security and Governance Leaders
Need to tell a regulator which human stood behind an agent's decision. They also need to decide what prompts and responses may appear in telemetry, and keep content capture off wherever they can.
Webinar Host

Atulpriya Sharma
Senior Developer Advocate, Improving
Atulpriya Sharma is a Senior Developer Advocate at Improving and a CNCF Ambassador. He has more than a decade of experience, moving from manual testing into cloud native technologies, with a focus on platform engineering and developer experience. He organizes the CNCF Platform Engineering Community, has spoken at KubeCon, KCDs and PlatformCon, and leads Cloud Native Hyderabad events.
