They Trace Calls. I Trace Decisions.

Ask a modern AI observability tool what your agents did today and you’ll get a beautiful answer: traces, spans, token counts, latencies. A whole genre of tooling has grown up to answer the question “what did the model do in this request?”

Which is a great question, if you’re a developer debugging a request.

I’m not. I’m an owner running a business. My questions are different: What was decided? On what evidence? At what cost? And was it right? Every dashboard I’d ever seen went quiet at exactly that point, so I built my telemetry to start there instead. The unit of trace isn’t the API call. It’s the decision.

Every proposed action in my system carries its whole causal chain: the dated data rows that motivated it, the reasoning that drafted it, the human who ruled on it, what actually executed, what was promised at birth, and the verdict that eventually landed. You can walk it in either direction, from a verdict back to its evidence, or from a piece of data forward to everything it ever caused. Call-tracing can’t do that, because calls don’t carry accountability. Decisions do.

The part I care about most is the cost accounting, because it counts the one thing nobody else does: attention. Compute is cheap and getting cheaper. My attention is not. So each decision accrues its model tokens, its tool time, and the seconds of human attention it consumed, which finally makes the real sentence possible: this decision cost this much, took this much of your attention, and returned this result. Efficiency for an operating company was never tokens per dollar. It’s outcomes per unit of founder attention.

And the honesty rules aren’t a values deck. Every number wears its data-date. Gaps never render as zeros. No verdict without a pre-registered plan. These are enforced as machine checks over the record, rules you can grep, not values you must trust. A discipline you can audit is a discipline. Anything else is a mood.

Where this leads is the fun part: seats with track records. Predictions at birth plus verdicts at death equals calibration over time, hit rates, cost per hit, per chair. The org chart becomes an instrument panel, and “should this agent get more autonomy?” stops being vibes and becomes a numbers review.

They trace calls. I keep score.

The full reference page, including how agent sessions get distilled into auditable profiles, lives at thinkfigaro.com/telemetry.