The $2,800 Wake-Up Call

My phone buzzed. Not a notification. An email from my cloud provider.

I sat up in bed. $2,793.00. In one hour. While I was asleep.

The email arrived at 3:14 AM. I read it at 6:17 AM. Three hours too late.

What Happened

An AI agent I'd deployed that afternoon had entered a retry loop. It was calling a premium LLM endpoint that cost $133 per call. Each retry consumed the same 200K-token context window. And it retried 21 times before the budget alert email even left the server.

The agent wasn't broken. It was doing exactly what I told it to do. The problem was that nobody told it how much money it was allowed to spend.

What I Checked First

I checked my API rate limits. They were fine — the provider was happy to take 21 calls at $133 each. Rate limits protect the provider, not me.

I checked my budget alert settings. They were set to trigger at $2,000. They triggered at 3:14 AM. The alert email arrived in my inbox. I just didn't see it until morning.

I checked my observability dashboard. It showed me exactly what happened: 21 calls, $133 each, starting at 3:01 AM. Beautiful graphs. Clear breakdown. Zero prevention.

The Realization

Every tool I had was reactive. They told me what happened AFTER the money was spent. None of them could stop the transaction BEFORE it executed.

Rate limits? They protect the API provider.

Budget alerts? They arrive by email.

Observability dashboards? They're beautiful tombstones for your wallet.

What I Built

I built AgentShield — a per-transaction spend firewall that evaluates every API call against configurable rules before it executes.

If that agent had been running AgentShield, the second call — the first retry — would have been blocked at $266. Not $2,793.

The evaluation takes less than 1 millisecond. Pure Python standard library. Zero dependencies. The agent gets a structured error. Your wallet stays closed.

The Offer

An unprotected runaway agent: $2,800
A whole year of AgentShield: $228