Give Your AI Investigator a Complete Case File (Sept. 9th)
Wednesday, September 9th, 2026: 11:00 AM to 12:00 PM
This session reframes the classic cost-versus-fidelity trade-off for the agentic era. When an AI investigator hits a sampled-out span, a rolled-up metric that no longer holds the dimension it needs, or a retention wall just short of the pattern it's chasing, it doesn't error out. It reaches a confident, wrong conclusion and you find out one incident at a time. We'll unpack that quiet failure mode and give you a vendor-neutral way to audit your own pipeline before you hand an agent the case.
Virtual
For over a decade, observability has been shaped by one fact: keeping telemetry is expensive, so teams throw most of it away. You sample traces. You cut retention to 7 or 15 days. You roll raw events up into aggregates and build pipelines to drop data before it ever lands. Every one of those is a decision made in advance about which questions will matter later. A fair bet for a dashboard a human configured, and a losing one for an AI investigator whose whole job is to ask the questions nobody planned for.
Key Takeaways:
1. Traditional observability optimizes storage around known queries; agentic observability has to survive the unknown ones, because the agent's job is to ask what no one anticipated.
2. Silent incompleteness is the real risk. An agent working from sampled, rolled-up, or aged-out data rarely errors out — it hands you a confident answer built on evidence that isn't there.
3. Retention, sampling, and rollups are decisions, not defaults. Each is a legitimate tool that also forecloses future questions. In an agentic world, those choices deserve to be made deliberately.
Hosted by DevOps.com