What we learned from 12 million runs
Patterns across a year of traces: where agents spend time, money and trust.
T
Maya Brandt
9 min
We looked at twelve months of anonymised, aggregate trace metadata. No prompts, no responses, only shapes.
Time
The median run spends 61% of its time waiting on tools, not models.
Money
Retries and oversized context windows account for a third of spend in teams without budgets.
Trust
Teams that gate deploys on evals roll back four times less often.
