Soft cap (alert)
- Status flips ok → alert at your threshold %
- Emits BudgetThresholdAlerted on cost.events
- Runs keep executing — the team gets a heads-up
- Watchable by the Real-Time Hub for automated response
Cost & value
Cap agent spend before it runs away.
Set rolling 30-day USD budgets per tenant and per agent, alert at a threshold you choose, and let a hard cap block runs the moment projected spend would exceed it — then prove the ROI in hours saved, cases resolved, and cost avoided.
402 BudgetExceeded · live spend · adoption & ROI
Cost Governance
$7,820
of $10,000 · 30-day cap
Hours saved1,240h
Cost avoided$312,000
budget ▸ hard cap blocks runs at 100% (402)
The problem it solves
Unmetered agents turn a token bill into a board problem.
An agent in a loop, a runaway batch, or a fleet quietly scaling all spend the same way: with no ceiling and no warning until the invoice lands. Cortex computes spend live from every run and gates the next one against a budget you set — so a $0.0001 cap returns 402 before the model is ever called.
How it works
A budget policy, evaluated live, enforced as a gate.
Every budget is a scope, a 30-day USD ceiling, an alert threshold, and a hard-cap switch. Spend is read live from actual run cost — no drift-prone counters — and the applicable tenant and agent policies are evaluated together before each run.
- 01
Set the policy
Scope to the whole tenant or a single agent. Set periodUsd (the rolling 30-day cap, 0 = unlimited), an alertThresholdPct, and whether the cap is hard. Admin-only.
- 02
Spend computed live
Cortex sums actual cost from skill runs over the trailing 30 days — tenant-wide, and per-agent via run metadata — and projects what the next run would add.
- 03
Gate every run
evaluateBudget returns { pct, projectedUsd, status, blocked }. checkBudgets combines tenant + agent policies and blocks if any hard cap would be exceeded — fail-closed, before the model is called.
Two ceilings, one runtime
Alert softly. Block hard. Always know.
Soft caps warn and keep working; hard caps stop the run cold. Both emit events the Real-Time Hub and Provenance can act on.
Hard cap (block)
- Run blocked once projected spend exceeds the cap (402)
- Emits BudgetExceeded with the blocking scope
- Fail-closed: deny is the default at the ceiling
- Tenant + agent caps combine — any hard breach stops it
Scope & period
- Per-tenant ceiling across all agents
- Per-agent ceiling for high-risk or high-volume work
- Rolling 30-day window — no calendar-month cliffs
- Live spend + status on the Cost Governance screen
Prove it — don't just claim it
Set a $0.0001 cap and watch the next run come back 402.
The budget gate is demoable end to end: a sub-cap run is allowed, the threshold trips an alert, and a run that would breach a hard cap is blocked before it ever reaches the model — with the blocking scope named in the response.
- Sub-cap run → allowed: true
- Threshold tripped → status: alert + BudgetThresholdAlerted
- Hard-cap breach → 402 BudgetExceeded, blockingScope: agent
- Status endpoint without admin → 401 (admin-gated)
budget-governance/check
policy ▸ scope=agent · periodUsd=$0.0001 · hardCap=true
Triage agentallowed: true
Triage agentstatus: alert
Triage agent402 BudgetExceeded
blockingScope: agent · run never reaches the model
POST /v1/skill-runs ▸ pre-run gate · fail-closed
From cost to value
The other half of the spend: what it bought you.
Cost governance is one side of the ledger. Cortex tracks the other — turning agent activity into the adoption and ROI metrics an executive actually reports on, every initiative inventoried, every outcome recorded.
Adoption & ROI
- Hours saved across automated work
- Cases resolved by agents vs handed off
- Revenue influenced and cost avoided
- Adoption and ROI dashboards by agent and team
AI portfolio
- Every agent, model, workflow, and prompt inventoried
- Business owner and risk tier on each initiative
- Spend and value attributed back to the portfolio
- One register from pilot to production
Work ledger
- Per-run cost recorded as the work happens
- The source of truth that drives live budgets
- Traceable to agent, run, and outcome
- Every initiative ties to business value + audit evidence
30-dayRolling spend window per policy
402Returned the instant a hard cap is breached
2Scopes that combine: tenant + agent
100%Live spend — read from real run cost
Works with the rest of the runtime
Budgets are one gate in a governed pipeline.
Cost governance shares the same identity, oversight, audit, and observability fabric as everything else in Cortex.
Control Tower
See live spend, alerts, and blocked runs across the fleet — and pause everything from one screen.
Agent IAM
Per-agent budgets are scoped to the identity — owner, risk tier, and the agent the spend belongs to.
Observability
Per-run cost shows up in run traces and quality scoring — the same data that drives live budgets.
Model Ops
Route to cheaper providers and enforce quotas upstream of the budget gate to keep spend in check.
Trust Ledger
Budget alerts and blocks are events on cost.events — surfaced in provenance and the audit trail.
Automation
The Real-Time Hub can watch BudgetThresholdAlerted and act — open an incident or tighten oversight.
Compliance framing
Spend control your auditors and your CFO both recognise.
Admin-gated policy, fail-closed hard caps, live cost attribution, and emitted budget events map to the cost-management and AI-oversight controls in the regimes you report against. Aligned with — never claiming certification you don't hold.
ISO 42001 · AI oversightNIST AI RMF · ManageSOC 2 · change controlEU AI Act · monitoringFinOps · cost attributionISO 27001 · admin access
“We set a hard cap per agent on day one. The first runaway batch hit the ceiling and came back 402 instead of a five-figure surprise — and finance finally had a number for what the agents were actually saving us.”
Put a ceiling on spend. Put a number on value.
Give every agent a budget that blocks before it overruns — and a ledger that proves what the work was worth.