Corey James- $0the cap on any day with no budget set
- 1/100 ¢the precision of per-second spend accounting
- 10 sbetween updates of per-container, per-schema and global spend
- 30 sbetween health passes that stop paid containers at the cap
I build spend control into the path that spends, and I keep its key in human hands. Rented capacity and metered APIs get hard gates; model usage inside governed work gets attribution.
A daily cap inside the provisioning path
Agents on FAFO Core and Function Forge can stand up GPU capacity, and a rented GPU bills from power-on whether anyone remembers it or not. I decided the spending limit would live inside the provisioning path, and that its only key would be a person at a terminal.
The limit is a daily dollar cap on the operator's calendar. A day runs from local midnight to local midnight in a timezone stored once, so every host computes the same day. Budgets are set per date or per date range, and every change lands in an audit table.
The cap is set with a command-line tool that only a person runs. That tool is the only code that writes budgets. No MCP tool on any gateway changes them, and nothing an agent passes to a provisioning call moves the cap.
The launch check
Every container launch passes the check before a port is leased or a cloud call is made, so a refusal leaves nothing to clean up and setting the budget is the daily switch for all automated provisioning. Work that runs in-process never touches the gate.
Inference Fabric GPU consumers, Agent Swarm persona containers and Helix render instances all pass this same gate. The persona launcher and the media controller treat any budget error as a refusal and spawn nothing. Admission compares actual spend with the cap, and the termination pass bounds any overshoot.
I kept projections out of the check because an exact rule is easier to trust.
Spend counted per second
Spend is accounted per second, starting the moment the cloud API confirms creation, because that is how the invoice is computed.
Spend is reported where it is spent. Terminating a node returns its final cost. The budget tool's status view shows the day's cap, spend and remaining budget, paid containers with their hourly rates, time to exhaustion, the next week's budgets and recent changes. Network-volume storage is synced from the provider and held to a monthly soft cap that only warns, because enforcing it would delete pre-warmed models.
A metered API with its own budget
FAFOHive's in-house X module spends a metered data API. It keeps a monthly unit budget and reserves units before each call, reconciling the reservations against the vendor's own usage figures.
The mark warns at 80 percent; the end of the bar refuses at 100.
Cost as telemetry in governed work
AgentOS attributes model usage to the work that spent it, taken from session transcripts and signed seat events. I ruled that cost inside governed work is telemetry and strategy, and removed a per-dispatch budget gate on quality-assurance work. The attribution design is on the Cost attribution for AI work page, and the provisioning path is on the Federated compute provisioning page.