Interactive · agent spend planner

What your agent costs you per month — and where each cap goes

Move the numbers to match your own setup. The two budgets that matter are inference (tokens burned thinking) and spending (things the agent buys). A cap on one does nothing for the other.

Cost per run — tokens, tool calls, retries

USD

Runs per day

Retry multiplier — sub-agents & loops

An agent that retries and fans out does not run one task — it runs many.

Fixed subscriptions — per month

USD

Purchases it makes for you — per month

USD

Cap	Where it must be enforced	What it protects
Inference ceiling	The provider console	The wall that stops a runaway loop from becoming an incident
Per-task estimate	Stated by the agent before the run	Surprises — a number in the transcript beats a trend on a dashboard
Velocity alert	The harness that dispatches work	Rate, not totals: a cap that trips on the 30th is a month of damage
Per-transaction cap	The card or wallet credential	One bad purchase — small enough to be annoying, not painful
Approval gate	Your own phone, above a threshold	Everything irreversible, and it makes the other four survivable

Your inputs, your arithmetic — nothing here is a benchmark. Set the inference ceiling at roughly four times a measured week plus 50% headroom, and write one rule into the agent: on reaching a limit, stop and report — never downgrade or improvise to finish.