Section 01 · OTel GenAI
One wire format, any backend.
every agent call · six numbers that matter
Record once. Read anywhere. The vendor becomes a query interface, not a cage.
Section 02 · cost attribution
$ per span, $ per agent.
7-day cost breakdown · one agent
Cost is a span attribute. Alert on rising cost like you alert on rising latency.
Section 03 · byte-deterministic replay
Reproduce the exact behavior.
scroll + seed = same output every time
Without deterministic replay, every agent bug is a one-time event. With it, every bug is a test case.
Section 04 · drift on the response
Watch distributions, not metrics.
distribution drift on output length
A single average won't catch this. A KS test on the distribution will, the day it emerges.
Section 05 · tools 2026
Where the OSS picks land.
OTel-native ↑↓ proprietary · OSS ←→ SaaS
Pick OTel-native + OSS. Add a hosted backend if your team prefers a UI. Avoid proprietary wire formats.
Section 05b · transcript + replay
Two implementations. Same bytes.
tamper-proof transcript + the exact recipe to run it again
A scroll proves what was said. A rerun-bundle proves it can be reproduced. Together: dispute-grade evidence.
Section 06 · vollko OSS
The primitives.
· · ·
Build the AI-native firm