Get started
Controls

Setting Spending Limits and Approvals for AI Agents

Learn how AgentWallet combines hard per-transaction, daily, and monthly limits with a human-in-the-loop approval policy to keep AI agent spending safe.

The AgentWallet TeamMay 24, 20262 min read
In short

AgentWallet controls agent spending two ways at once. Hard limits set per-transaction, daily, and monthly caps that block any action that would exceed them. A soft approval policy lets in-limit actions proceed only after a human reviews and approves them, keeping a person in the loop for anything sensitive.

Hard limits: the caps that cannot be crossed

Per-agent policy sets hard spending limits across three windows: per-transaction, per-day, and per-month. An action that would push the agent past any of these is blocked outright. These caps are terminal, meaning no approval can override them.

Soft approvals: keeping a human in the loop

Under the hard caps sits a soft approval policy. When an action is within the hard limits but crosses the approval policy, AgentWallet does not execute it. Instead it creates an approval request. A human reviews the request and, only when they click approve, the action runs and moves money server-side.

  1. The agent attempts an action.
  2. AgentWallet checks it against the hard limits; if over, it is blocked.
  3. If within limits but over policy, an approval request is created.
  4. A human approves, and only that click moves the money.

Enforced atomically under load

Velocity and balance are enforced atomically under a per-agent lock, so bursts of concurrent agent activity cannot slip past a cap. The number you set is the number that holds.

Frequently asked questions

Can an approval override a hard limit?
No. Hard limits are terminal. Approvals only apply to actions that are within the hard limits but cross the soft approval policy.
Who can approve an agent's action?
A human reviewer in the dashboard. The money only moves at the moment they click approve.
What spending windows can I cap?
Per-transaction, daily, and monthly limits, each enforced independently.

Sources

  1. Model Context Protocol Anthropic

Give your agents a wallet

Per-agent wallets, virtual cards, USDC on Base, and policy guardrails — provisioned through one MCP endpoint.

Start free