Play · Developers and platform teams
Cap agent spend before finance notices
A budget thatstops the runat the limit.
Observe prices every call across providers in one dollar figure, per agent, per customer, per feature. A budget is a policy: when a run reaches it, the run stops and the record says why.
30 minutes. No commitment.
SOC 2 Type II attested · ISO/IEC 27001:2022 certified · HIPAA with a BAA · aligned to the EU AI Act and NIST AI RMF
Who
A platform team running a research agent fleet across Anthropic, OpenAI and Bedrock, with one shared company card.
What happened
A retry loop ran for a weekend and the month's model spend tripled. Finance found out from the invoice.
What's at stake
Spend that is only visible after the fact, no way to say which agent or customer caused it, and nothing that stops a loop before it finishes the bill.
- 01Observe
One dollar figure across providers.
Two lines, an import and waxell.init(), instrument the fleet; 160+ instrumentors cover Anthropic, OpenAI, Bedrock, Vertex and the frameworks on top of them. Every call lands with tokens, latency, and cost in real dollars at the price you pay. The executions list shows spend per agent, per customer, and per feature, so the weekend loop is a row you can open, not a line on an invoice.
waxell.init() · 160+ instrumentors · cost per run in real dollars · per agent, per customer, per feature
observe · executions
- 02Enforce
Set the budget. The run stops.
Set a limit per agent, per model, per end user, or across the system; nested limits resolve to the tightest. When a run reaches it, the run stops in BLOCKED, a final state, and the record names the budget that stopped it. Observe also suggests step limits and timeouts from run history, so the loop that caused the bill gets a ceiling as well.
per agent · per model · per end user · system-wide · run stops, BLOCKED · recommendations from run history
govern · budgets
- 03Prove
Finance gets a number per agent.
Every spend decision carries the budget that made it, the agent, and the customer or feature it was for. The executive view shows spend against limits over the period, and wax budgets show answers the same question from the terminal.
wax budgets show · spend against limit per agent · durable, exportable decision record
insights · executive
- anthropic.messages · 41k tok · $0.62research-fleet-03 · budget 62% usedallow
- anthropic.messages · 58k tokresearch-fleet-03 · budget reached · run BLOCKED at step 14block
- openai.chat · 12k tok · $0.09summarizer · customer acme · end-user budget 31% usedallow
- bedrock.converse · retry 40crawler-02 · step 212 · step limit 200 · from recommendationblock
Illustrative entries in the shape the product writes them.
Products in this play
Dollars. One figure across providers, converted at the price you actually pay. Tokens in and out are still on every call for the engineers who want them.
Run this play this week.
Free for your first agents and devices, live in about thirty seconds. Or walk through it with someone who has run it before.
30 minutes. No commitment.
SOC 2 Type II attested · ISO/IEC 27001:2022 certified · HIPAA with a BAA · aligned to the EU AI Act and NIST AI RMF