Spend controls for AI agents, in practice.
Why alerts and prompts don't stop an agent from overspending, and how to enforce a per-call and daily cap in code, with a tested example.
Three ways to put a human in front of MCP tools, which one an agent can't skip, and a tested setup for Claude Desktop and other MCP hosts.
What the PocketOS and DataTalks.Club incidents teach, which safeguards would have mattered, and what a tool policy can and can't prevent.
Let an agent handle small refunds and send big ones to a human, using PayPal's MCP server, a dollar limit and a daily budget.
Allow internal email, hold everything else for a human, and check every recipient before an agent's message leaves your code.
We edited and deleted records in an agent's audit log to see what gets caught. Why a hash chain isn't enough on its own, and what to add.
A plain definition: the five controls that limit what an agent can spend, where they should run, and how they differ from alerts and prompts.
How four tools put a budget on x402 payments: who holds the key, where the decision runs, what a human can stop, and what survives a restart.