Track AI agent token spend per task

Where the numbers live

Every model call the agent makes is logged with its token count and cost. The audit history is the raw record; the console’s insights view is the same data, pre-aggregated.

Build the per-task picture

1. Open the audit export and group rows by task id.
2. Sum tokens and cost per task.
3. Sort by cost and take the top five — those are your repeat offenders.
4. Compare: what did each expensive task actually deliver?

What the patterns mean

• One task that costs more than all others: check whether it is looping — its Deep Dive trace will show repeated near-identical steps.
• Steady growth across all tasks: your prompts are getting longer — compact context instead.
• Spikes at the same time every day: a scheduled job you forgot about.

Turn tracking into limits

Tracking answers “where did it go”; token budgets and spend alerts turn that into a hard stop. The combination — measure, then limit — is the whole of cost control.