AI agent cost control: budget caps and credit gates without runaway spend
Every agent build calls a live model, so cost control is built into the run itself rather than split into a separate mode.
Per-run budget caps
Every run carries a budget cap. The runtime ledger accumulates cost step by step and stops a run the moment it exceeds its cap — a runaway loop never becomes an invoice.
Evaluation has its own execution and judge budgets. Both can be capped from the compile settings; leaving them empty applies the runtime defaults.
The credit gate
Runs draw down prepaid credits. A gate checks the balance before execution, so exhausted credits stop further provider calls. In beta, provider billing rules and London System Agent's prepaid-credit policy both matter.
Cost visibility
After a run, the cost panel shows per-step input/output tokens and spend. The pricing table uses approximate public list prices, so treat it as an operational meter and optimization aid, not a final invoice ledger.
Your turn to build
Compile your first agent graph from a single sentence.