Four AI Agents, One Infinite Loop, $47,000 in 11 Days

A 4-agent LangChain system was left to hand tasks between itself, one agent to the next, with no exit condition. It ran unattended for 11 days. The bill was $47,000. Nobody noticed until it was over.

What Happened

A multi-agent system built on LangChain was set up so that four agents could hand work off to one another — agent A finishes a step and passes context to agent B, B to C, C to D, and back around. It's a common pattern for breaking a complex task into specialised roles. The problem is what happens when the handoff loop has no exit condition and no budget ceiling: the agents kept the conversation going with each other, each round-trip triggering fresh LLM calls, for eleven straight days. No human was reviewing the session. No cost alert fired in time to matter. By the time anyone looked at the bill, it read $47,000.

This isn't a hypothetical. It's one of a growing pattern of "agent left running, nobody checked the meter" incidents now common enough that vendors are shipping dedicated "token budget enforcement" products specifically to stop it (see Sources below).

Why This Keeps Happening

The failure isn't the model, and it isn't malice. It's a structural gap in how agent loops get written:

• A while or agent-to-agent handoff loop gets written to keep going "until the task is done" — with no hard iteration cap as a backstop.
• Each LLM call inside the loop has no output token ceiling, so a single expensive round can cost far more than expected.
• Cost alerts, where they exist at all, are usually daily or weekly digests — not real-time circuit breakers. By the time an alert lands, days of spend have already happened.
• Multi-agent systems compound the problem: four agents talking to each other means four times the call volume per "round", so a loop that would be merely wasteful with one agent becomes financially serious with four.

// no iteration cap, no budget guard, no token ceiling
while (!taskComplete) {
  const response = await llm.chat.completions.create({
    model: 'gpt-4',
    messages: conversationHistory,
  });
  conversationHistory.push(response);
  taskComplete = checkIfDone(response); // agent decides for itself when to stop
}

How Thuban Cost Guard Catches This

This is precisely what TCG-101 (Unbounded AI Loop Execution), one of Cost Guard's two v0.1 detection rules, is built for. Cost Guard's static analyzer walks every while, do-while, and for loop in your codebase and checks whether an AI SDK call inside it (OpenAI, Anthropic, Vercel AI SDK) has a genuine bounded exit — a real iteration counter, a budget-guard wrapper, or a reachable break condition. A loop whose only exit condition is "the model decides it's done" doesn't count. That's flagged CRITICAL, before the code ever ships, in every thuban scan.

Sources

• Waxell: AI Agent Token Budget Enforcement — Why Alerts Fail and What Actually Works
• LinkedIn: An AI agent system burned $47,000 in API costs

npx thuban protect