How is this different from LLM observability tools?

Call-tracing tools answer "what did the model do in this request" — spans, tokens, latencies. Useful for developers, silent on the question a business owner actually has: what was decided, based on what, and did it work. Figaro roots its trace at the decision. Every proposed action links the data that motivated it, the reasoning that drafted it, the human who ruled on it, what executed, what it cost, and the verdict it earned. The unit of accountability is the judgment, not the API call.

What does a decision’s audit trail contain?

The chain runs both directions: from any verdict you can walk back to the exact dated data rows the seat sensed, the transcript that drafted the proposal, who approved it and how long they spent, and the external reference of what actually executed. From any piece of data you can walk forward to everything it ever caused. Because the underlying record is append-only, the chain cannot be quietly edited after the fact.

What does a decision cost, and how do you know?

Each decision accrues its compute (model tokens), its tool time, and — the number no one else tracks — the human attention it consumed, measured from the approval surface. That makes the real question answerable per seat and per brand: what does a judgment cost here, and what did it return? Efficiency for an operating company is not tokens per dollar; it is outcomes per unit of founder attention.

How do you audit what an agent did in a session?

Every headless agent run is captured and distilled into a profile: which tools it used and how often, what it read and wrote, what it fetched, what was denied, what it cost. Deterministic rules flag the things an owner should see — unexpected installs, out-of-scope writes, unfamiliar network calls. Most audit questions are answered from profiles without anyone reading a transcript; the transcript is there when you need the whole story.

Can seats be scored over time?

That is where this design leads: every action is born with a predicted impact and dies with a measured verdict, so each seat accrues a track record — hit rate, calibration, cost per hit, time-to-decision. Autonomy promotions stop being vibes and become a review of the numbers. We publish the principle and keep our evaluation suites and calibration data private, for the same reason a good team publishes its philosophy and not its playbook.

Reference · Observability

Telemetry: We Trace Decisions, Not Calls

Generic AI observability records what a model did in a request. Figaro records what a company decided, on what evidence, at what cost, and whether it was right.

Published August 9, 2026

There is a generation of AI observability tooling built for developers: traces of model calls, token counts, latency spans. All useful, all beside the point for an operating company. The owner’s questions are different: what was decided, on what evidence, at what cost — and was it right? Figaro’s telemetry is built to answer those.

The decision is the unit of trace

Every proposed action carries its causal chain: the dated data rows the seat sensed, the reasoning that drafted it, who proposed, who decided, what executed, what was promised at birth, and the verdict that landed. One query walks 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 say that, because calls don’t carry accountability. Decisions do.

Cost per judgment, including the human

Compute is cheap and getting cheaper; founder attention is not. So each decision accrues all three costs — model tokens, tool time, and the seconds of human attention it consumed — and rolls them up per seat, per loop, per brand. The sentence this makes possible is the one that matters: this decision cost this much, took this much of your attention, and returned this result.

Rules you can grep, not values you must trust

The honesty rails — every number wears its data-date, gaps never render as zeros, no verdict without a pre-registered plan, no action without a gate approval — are enforced as machine checks over the record, not promised in a values deck. A discipline you can audit is a discipline; anything else is a mood.

Agent sessions, distilled

Headless agent runs are captured and boiled down to profiles — tools used, files touched, network reached, denials hit, cost incurred — with deterministic flags for what an owner should see. Sunlight over agent behavior is the anti-abuse mechanism that scales: an agent doing sloppy or weird work is visible doing it.

Where it leads: seats with track records

Predictions at birth plus verdicts at death equals calibration over time: hit rates, cost per hit, time-to-decision, per chair. The org chart becomes an instrument panel, and autonomy promotion becomes a numbers review. The principle is public; our evaluation suites and calibration data are not — publish the philosophy, keep the playbook.

Questions founders ask

How is this different from LLM observability tools?: Call-tracing tools answer "what did the model do in this request" — spans, tokens, latencies. Useful for developers, silent on the question a business owner actually has: what was decided, based on what, and did it work. Figaro roots its trace at the decision. Every proposed action links the data that motivated it, the reasoning that drafted it, the human who ruled on it, what executed, what it cost, and the verdict it earned. The unit of accountability is the judgment, not the API call.
What does a decision’s audit trail contain?: The chain runs both directions: from any verdict you can walk back to the exact dated data rows the seat sensed, the transcript that drafted the proposal, who approved it and how long they spent, and the external reference of what actually executed. From any piece of data you can walk forward to everything it ever caused. Because the underlying record is append-only, the chain cannot be quietly edited after the fact.
What does a decision cost, and how do you know?: Each decision accrues its compute (model tokens), its tool time, and — the number no one else tracks — the human attention it consumed, measured from the approval surface. That makes the real question answerable per seat and per brand: what does a judgment cost here, and what did it return? Efficiency for an operating company is not tokens per dollar; it is outcomes per unit of founder attention.
How do you audit what an agent did in a session?: Every headless agent run is captured and distilled into a profile: which tools it used and how often, what it read and wrote, what it fetched, what was denied, what it cost. Deterministic rules flag the things an owner should see — unexpected installs, out-of-scope writes, unfamiliar network calls. Most audit questions are answered from profiles without anyone reading a transcript; the transcript is there when you need the whole story.
Can seats be scored over time?: That is where this design leads: every action is born with a predicted impact and dies with a measured verdict, so each seat accrues a track record — hit rate, calibration, cost per hit, time-to-decision. Autonomy promotions stop being vibes and become a review of the numbers. We publish the principle and keep our evaluation suites and calibration data private, for the same reason a good team publishes its philosophy and not its playbook.

Drafted by the Figaro content seat · edited by Fable · reviewed by Kyle · last updated August 9, 2026