Cost Tracking

How cost attribution works

Every model call, tool invocation, and compute operation is tagged with the agent, team, task, and user that initiated it — the same registry identity used everywhere else in Prefactor. A simple question might cost 500 tokens; the same agent handling a multi-step research task might run 50,000 tokens across several model calls and three tool invocations. Attribution captures that variability at the task level, in real time, instead of averaging it away in a monthly rollup.

Cost dashboards drill down from organisation to team to agent to task, so a spend spike can be traced back to the specific run that caused it rather than showing up only as a line on next month's invoice.

Budget enforcement at runtime

Cost limits are expressed the same way as any other Prefactor policy — declaratively, and enforced at the point an agent tries to act. Teams set daily, weekly, or monthly budgets per agent or per team; when an agent approaches its limit, Prefactor can throttle it or block further calls outright, rather than letting spend run past the limit before anyone notices.

Because enforcement happens at runtime, a budget breach is stopped mid-task instead of being caught in a billing review days later.

Anomaly detection and alerting

Cost tracking also watches for spending patterns that don't fit an agent's normal profile — a loop that keeps calling the same tool, a model upgrade that quietly changed the per-call price, or a misconfigured agent burning through its budget in an hour instead of a month. Anomalies route to Slack, PagerDuty, or the alert channel your team already watches, so the right people find out while the spend is still small enough to matter.