The 4-Layer Policy Cascade That Bounds AI Agent Spending
AgentWallet enforces spending controls that cascade from account to principal to agent, where each level can only tighten the one above it. Here is how.
AgentWallet bounds agent spending with a four-layer policy cascade: the account sets the outer limit, each principal narrows it, and each agent narrows it again. A child level can only ever be more restrictive than its parent, so authority always shrinks as it flows down to the agent that actually spends.
What cascades, exactly
Policy covers hard spending limits and a soft approval policy. Hard limits are per-transaction, daily, and monthly caps. The soft approval policy decides which actions need a human sign-off before they run. Both are inherited: an agent starts from its principal's envelope, and the principal starts from the account's.
Why narrowing-only matters
The narrowing-only rule is what makes delegation safe at scale. Because no lower level can widen what a higher level set, you can hand an agent to a team or to another automated system without worrying that a local change quietly grants it more spending power than intended. Least privilege becomes the default rather than something you have to police.
| Level | Represents | Typical setting |
|---|---|---|
| Account | The business tenant | Outer ceiling for all spending |
| Principal | A person or budget | A share of the account ceiling |
| Agent | The AI actor | The task-specific limit it runs under |
Hard limits versus approvals
- Over a hard limit: the action is blocked and never executes.
- Over the approval policy but within hard limits: a human approval request is created.
- Within policy: the action proceeds and is recorded.
Enforced atomically
Velocity and balance are checked atomically under a per-agent lock, so two concurrent agent actions cannot race past a cap at the same moment. The limit you configure is the limit that holds, even under bursty automated traffic.
Frequently asked questions
- Can I give one agent more room than its principal?
- No. An agent can only ever be equal to or more restrictive than its principal, which in turn cannot exceed the account.
- What is the difference between a hard limit and the approval policy?
- A hard limit blocks an over-cap action entirely. The approval policy lets an in-limit action proceed only after a human approves it.
- Do concurrent actions risk double-spending past a cap?
- No. Velocity and balance are enforced atomically under a per-agent lock so concurrent actions cannot race past a limit.
Sources
Give your agents a wallet
Per-agent wallets, virtual cards, USDC on Base, and policy guardrails — provisioned through one MCP endpoint.
Start free