Product
Cost Guards
A hard ceiling per run. The agent stops at the limit instead of quietly overspending.
Ceilings apply per invocation, per agent, or per workspace, and count model tokens, tool-side credits, and retries together. A run that stops on budget is marked distinctly from one that failed, because they need different responses.
Warehouse credits count too
An expensive analytical query is spend, not a free call, and the ceiling treats it that way.
Alerts before the wall
A notification fires at eighty percent so somebody can raise the ceiling deliberately.
Attributed per step
You can see which step is expensive, not just that the run was.
budget: {
perRun: { usd: 0.40 },
alertAt: 0.8,
onExceed: "stop"
}Measured outcome
A prompt regression tripled token use overnight. The guard stopped 4,100 runs at their ceiling instead of producing a five-figure invoice.
Product
- Run GraphWatch every step, branch, and retry as it happens, with the exact inputs and outputs on each node.
- Retry PoliciesPer-tool exponential backoff, declared in one file instead of scattered through your handlers.
- Audit TrailEvery prompt, tool call, and token — retained and replayable, exportable as compliance evidence.