What's this about?
Operational telemetry is overwhelmingly engineered for human reading: systems repeatedly serialize verbose prose, static keys and redundant context across billions of log lines. As autonomous agents become the primary consumers of that output, this paper proposes emitting verifiable state-delta evidence instead.
What exactly is a "state delta"?
A log line typically describes an event in prose — "user 12345 successfully updated their billing address at 14:32." A state delta records only what changed: this field, from this value to this value, at this time. It's smaller, unambiguous, and machine-checkable. The reason it matters now is context economics: an agent reading logs pays per token, and conventional logs are enormously redundant — the same keys and boilerplate repeated billions of times. "Verifiable" is the second half: the evidence can be checked for tampering, so an agent's account of what happened can be audited.
Why is this worth noticing?
- It's the same realization arriving across the stack this month: tooling built for human ergonomics is wasteful when the consumer is a model paying per token — the same premise behind Pi's "build context-efficient tools."
- The verifiability angle connects to Mandato's chained audit trails: as agents operate systems, evidence of what actually changed becomes the accountability substrate.