Agent usage enforcement UsageFlow

Enforce usage limits on AI agents before the next step runs.

UsageFlow enforces per-customer token and spend limits on AI agents. Before each step, a policy reads the customer's usage and switches to a cheaper model, caps the budget, or blocks the call. No redeploy.

Get Early Access Read the docs

Free: 1 workflow with live policy. No credit card.

Why agent budgets fail when they live in code

A hard-coded step limit doesn't know which customer is running, what plan they are on, or what they already spent today. Changing it means a deploy, and every agent ends up with its own copy of the rule.

Budget in code	Budget in a UsageFlow policy
Same limit for every customer	Limit per customer and plan
Counts steps, not tokens or spend	Reads the customer's running usage
Fails the run when hit	Can downgrade first, then block
Changing it needs a deploy	Edit in the Console; applies to the next call

Per-customer limits for agent loops

Retries and loops are counted because every model call goes through the SDK with an identity and a workflow. The call that crosses a threshold still runs; the next one is checked against it.

Downgrade instead of failing

Set a ladder: at one threshold, route to a cheaper model; at the cap, block. Customers keep getting answers while spend stays bounded. Traces show the requested model next to the executed one.

Change limits without redeploying

Thresholds, models and who they apply to live in the UsageFlow Console and apply from the next call.

Usage enforcement vs. security guardrails

Both act before a call runs, but they answer different questions.

	Usage enforcement	Security guardrails
Question	May this customer spend more?	May the agent use this tool or data?
Inputs	Identity, plan, current usage	Prompt, tool call, data access
Actions	Route, cap, block	Allow or deny a permission
UsageFlow	Yes	No — use a dedicated guardrail tool

SDK example

Catch the rejection and end the run with a clear message. The provider is never contacted for a blocked call.

TypeScript

import { usage, VibeRejectionError } from '@usageflow/vibe';

try {
  const turn = await usage.chat({
    identity: customerId,          // whose budget this step spends
    workflow: 'research-agent',    // the policy that sets the limit
    provider: 'anthropic',
    model: 'claude-sonnet-5',
    messages,
  });
} catch (err) {
  if (err instanceof VibeRejectionError) {
    return 'This run reached your usage limit.';
  }
  throw err;
}

Keep going

Agent metering What each agent run consumes, per customer. AI usage metering Tokens, requests and cost per customer.

Agent usage enforcement FAQ

How do I enforce a token budget on an AI agent?

Pass the customer identity with each agent call. A UsageFlow policy checks that customer's current usage before the call runs and can route to a cheaper model, cap the budget, or block it.

Is this a security guardrail?

No. UsageFlow enforces usage and cost policies — budgets, quotas and model access by plan. Tool-permission and data-security guardrails are a different category.

Put a budget on every agent run

Free: 1 workflow with live policy. No credit card. TypeScript, Python and Go, with OpenAI and Anthropic.

Get Early Access Read the docs