Hard Spend Caps Are Becoming the Default. Where Should Agent Spend Be Stopped?
Vendors have started shipping spend controls that actually stop usage, and the argument that hard caps should be the default is getting wider attention. That is good news for anyone running an agent that pays for things. It is also easy to over-read. This post separates three controls that often get lumped together, explains what a provider cap cannot see, and ends with a short audit you can run this week.
What changed in October
Two items made the rounds this month. Simon Willison wrote on October 3 about spend limits and why hard caps, rather than after-the-fact alerts, should be the default for services that bill by usage; it also drew a large discussion on Hacker News. On October 8, Google Cloud announced a project-level spend cap. We are paraphrasing both from memory of their public posts, so read the originals and the current Google Cloud documentation for exact behaviour, scope, and any delay between hitting a cap and usage stopping.
The direction is clear even if the details vary by provider: enforcement is replacing reporting as the expected default.
Alert, cap, and gate are three different controls
These get confused, and the difference matters during an incident.
An alert describes spend. It tells a person something crossed a threshold. Nothing stops. Microsoft's documentation for Azure budgets, as we read it, describes thresholds that trigger notifications or actions you configure, and says a budget on its own does not stop resource consumption. Check the current Azure docs before relying on this, but it is a good example of a control that is real, useful, and not a limit.
A cap stops usage on one account or project. Once the limit is reached, the provider refuses further billable use in that scope. This is enforcement, but the scope is whatever the provider chose: an account, a project, sometimes a key. Providers also differ on whether a cap is hard or best-effort, so confirm which one you have.
A gate decides per transaction, before the call. The caller asks whether this specific payment or paid call is allowed, and proceeds only on a yes. The decision can consider the agent, the amount, the merchant, and what has already been spent.
Why provider caps are not enough for agents
A cap protects your invoice with that provider. For agents, that leaves gaps.
- Scope is coarse. Caps are typically per project, account, or key. They are mostly not per agent, per merchant, or per payment. A cap cannot tell you which of ten agents used the budget, or give one agent a smaller allowance than another.
- They stop at one provider. An agent that calls several paid services is bounded by several unrelated caps, each configured separately, and none of them knows what the agent spent elsewhere.
- They are often off by default. In most cases someone has to turn a cap on, so the configuration has to exist before the incident, not after it.
- They say nothing about why. A cap refuses usage. It does not record which rule applied to a given call or route an unusual one to a person.
None of this makes caps a bad idea. Keep them as the outer wall.
Where the check belongs
Put the decision in the request path, before the call, and log the reason with it. The agent, or the runtime that orchestrates it, asks for a decision, gets back approve, deny, or route to human approval, and only pays on approve. Because the decision is made outside the model, the agent cannot talk its way past it.
Paitify is one example of a layer that does this. It evaluates a spend request against per-agent policies and returns a decision with the rule behind it. It is a policy and audit layer, not a payment path: it does not move, hold, or custody funds, and it does not sit between your agent and the vendor. Your integration is responsible for asking first, and a check only covers calls that go through it.
For the approval-routing side of this, see Designing Auditable Guardrails.
A five-minute audit
Provider caps to confirm:
- For each paid provider your agents use, is a cap or hard limit actually enabled, not just an alert?
- Is the cap hard (usage stops) or soft (notification only)? Check the provider's current documentation.
- What is the scope: account, project, or key? Which of your agents share it?
- Who is notified when it trips, and what is the plan if production traffic stops?
Per-agent limits to add:
- A per-transaction ceiling for each agent.
- A daily or monthly budget for each agent, separate from the provider cap.
- An allowed list of merchants or endpoints the agent is expected to use.
- A decision for over-threshold requests: deny, or send to a person with an expiry.
- A decision log you can read, so you can see what was refused and why.
Try it on one agent
Pick one agent, put its paid calls behind an authorize step, and read the decision log for a week. The free tier includes one agent, one policy, and 1,000 authorizations per month. See the API reference for the authorize call, or watch the demo video on the homepage to see an agent approved and then stopped.