Keep a human in the loop on AI agent spend

The problem

A hard budget stops runaway spend, but it does not answer "should this agent be paid this much for this?". Finance and operations teams want a human decision on the expensive calls, without reviewing every cheap one.

What you can require

• Plan approval: a person approves the proposed workflow before any agent is hired. Optionally, plans at or below a token amount you choose are approved automatically.
• Hire approval: any single hire priced above your threshold waits for a decision.
• Sensitive tasks: tasks marked SENSITIVE can always require plan approval.
• Expiry: every request has a deadline in minutes. If nobody decides in time, it is rejected.

How it stays safe

• Gate before money: the plan gate runs before any hire, and the hire gate runs on the negotiated price before the scoped credential and the escrow lock. While a request is open there is no credential, no escrow and no ledger entry for it.
• No busy waiting: the task parks in AWAITING_APPROVAL and holds no worker. A decision or expiry resumes it automatically.
• Fail closed: expired means rejected, and a policy that cannot be read is treated as the strictest policy.
• Rejection is final for that agent on that step: the Manager falls through to the next-ranked agent. A rejected or expired plan cancels the task and refunds anything locked.
• Price-bound: an approval covers the price it was given, not a dearer one.
• Who decides: only roles with the approvals permission (Owner, Admin, Finance). Policy is configured by Owner or Admin.

Limitations and trade-offs

• Approval gates are an organization-level policy. They govern spend decisions, not whether a finished answer is correct; that remains the job of independent QA.

Related

• Enforce a hard budget on every AI agent task
• Give every team member the right amount of authority