Agent Spend Needs an Outcome Denominator

GitHub has moved per-user AI credit budgets into the billing interface, while its latest VS Code release makes agent usage and cost easier to inspect alongside parallel sessions. That is a useful market signal: agent spend is leaving the infrastructure footnotes and becoming an everyday operating decision. The control is necessary, but a budget without an outcome measure only tells a team how quickly it is consuming credits—not whether the work is valuable.

A flat cap can punish the people using agents for high-value work and quietly subsidise loops that generate activity without accepted output. Teams need a denominator. Cost per session is interesting; cost per verified outcome is actionable. Once quality and spend share the same view, product builders can decide which workflows to automate, redesign, route to a different model, or stop.

Instrument one repeatable agent workflow as a small unit-economics system. Trigger: a real job such as resolving a support issue, preparing a release, or shipping a scoped code change. Cost: model credits, runtime, tool calls, and human review minutes. Outcome: one binary acceptance event plus a quality score. Segments: workflow type, model, operator, and failure reason. Verifier: the artifact passes the same test a human-produced result would face. Decision rule: expand the workflow only when cost per accepted outcome improves without quality falling. Stop condition: pause any route that gets cheaper by moving hidden work into review or rework.

Take the last 20 runs of one agent workflow. Divide total AI spend and review time by the number of outputs that were actually accepted. Find the most expensive failure mode and remove that before raising the budget.