We Spent $30+ on an AI Agent Doing Nothing. Here’s What Happened.
Over two days, we spent more than $30 in AI model usage on an agent that was not completing a real task — just waking up every 30 minutes and burning tokens. Here is what happened and what we learned.
AI agents are powerful. They can read files, update their own instructions, run scheduled tasks, connect to messengers, and automate real work. But that also means one thing: if you do not monitor them carefully during setup, they can quietly burn money while doing absolutely nothing useful.
We learned this the hard way. Over two days, we spent more than $30 in AI model usage on an agent that was not completing a real task, not helping a customer, not researching anything important, and not producing business value. It was simply waking up, checking a file, trying to run something that did not exist, and spending tokens every time.
The setup
We were setting up an AI agent using OpenClaw — an open-source AI agent framework that can connect to messengers like Telegram and run with your own AI model API keys. After installation, we spent time configuring and training the agent: adjusting its instructions, testing how it responds, improving its behavior, checking its workspace files, and teaching it how to operate in our environment.
This is normal during early agent setup. You do not install an AI assistant once and immediately get a perfect business automation system. You configure it, test it, correct it, and gradually give it more responsibility. But during this process, something unexpected happened.
The hidden problem: HEARTBEAT.md
OpenClaw agents can use a file called HEARTBEAT.md. The basic idea is simple: the agent wakes up on a schedule, checks this file, and decides whether there is something it should do. In our case, the agent was waking up every 30 minutes.
At some point during setup, the agent apparently updated its own HEARTBEAT.md file and added one line. That line did not represent a real task. It did not point to a useful workflow. It did not describe a meaningful automation. But the agent still interpreted it as something to process.
So every 30 minutes, it woke up, checked the file, tried to start a process that did not exist, spent tokens, and then repeated the same cycle again. No visible output. No completed task. No business value. Just token usage.
Why this matters
This is a small example, but it points to a much bigger issue. AI agents are not simple chatbots. They can have memory, scheduled tasks, tool access, file access, self-updated instructions, long-running workflows, and integrations with external systems. That is what makes them useful.
But it also means they need to be treated like real software systems. If an agent has permission to update its own files, schedule work, or run background checks, small mistakes can create recurring cost. Not because the agent is bad. Not because the tool is broken. But because autonomous systems need monitoring, limits, and review.
The real cost was not just $30
Spending $30 was not the biggest issue. The real issue was that the agent was spending money silently. If we had not checked the workspace files and usage, the cost could have continued.
Now imagine the same pattern with a more expensive model, multiple agents, hourly or 30-minute wakeups, long system prompts, large context files, connected tools, and a business account with higher billing limits. The bill can grow quickly. This is especially important in the first weeks of setting up an AI assistant, when the system is still being trained, tested, and corrected.
What we learned
After every training or configuration session, review the agent’s workspace files — especially files that control behavior, scheduling, or automation.
For OpenClaw and similar agent systems, check files such as HEARTBEAT.md, agent instruction files, memory files, tool configuration files, automation or schedule files, and anything the agent is allowed to modify. Do not assume that because the agent seems quiet, it is inactive. An agent can be quiet in chat but still active in the background.
Practical checklist after setting up an AI agent
If you are configuring an AI assistant for yourself or your business, use this checklist.
1. Check recurring task files
Look for anything that causes the agent to wake up automatically — heartbeat files, cron schedules, background task definitions, recurring reminders, or autonomous workflow instructions. If you do not understand why a recurring task exists, disable it until you do.
2. Monitor model usage daily during the first week
Do not wait for the monthly bill. During the first week, check usage every day. Look for unexpected token spikes, activity during hours when you did not use the agent, repeated calls every 30 or 60 minutes, high usage from simple tasks, and usage from agents that should be idle.
3. Set billing limits
If you are using your own API keys, configure spending limits. This is especially important for OpenAI, Anthropic, OpenRouter, and any provider where usage is billed per token. Start with a conservative limit and increase it only after you understand the agent’s behavior.
4. Disable autonomous behavior at the beginning
In the first days, the agent should mostly respond when you ask it to do something. Avoid enabling too much autonomy too early. Be careful with wakeups every 30 minutes, automatic research, automatic email processing, self-created tasks, tool use without approval, and background workflows. Add autonomy gradually.
5. Review files after every training session
If you spent time teaching or configuring the agent, inspect what changed — especially if the agent can edit files. Check instructions, memory, heartbeat files, task files, and configuration files. Small changes can have recurring effects.
6. Keep approval rules clear
An AI assistant should know what it can do freely and what requires approval. For business use, we recommend requiring explicit confirmation before the agent sends emails, posts online, changes CRM data, deletes files, modifies shared documents, creates calendar events, runs infrastructure changes, or performs financial actions. Clear rules reduce risk.
AI agents are still worth it
This story is not an argument against AI agents — it is the opposite. AI agents are becoming a normal part of business operations. Every company will eventually need some version of an AI assistant for research, sales, operations, internal knowledge, support, reporting, or technical work. But they need to be implemented carefully.
An AI assistant is not just a prompt. It is a system. And systems need monitoring, permissions, usage limits, review processes, safe defaults, and clear ownership. If you treat an AI agent like a toy, you may get toy-level reliability. If you treat it like business infrastructure, it can become extremely useful.
Our recommendation
If you are testing AI agents locally, start small. Do not connect everything on day one. Do not enable background autonomy before you understand how the agent behaves. Watch your usage closely. Review the files the agent can edit.
And if you want an AI assistant for real business workflows, consider setting it up in a managed environment with proper monitoring, access control, and approval rules. That is exactly what we help companies do at Evolution AI — we design and deploy AI assistant environments that can support real work: research, sales, operations, CRM, Google Workspace, internal knowledge, and technical workflows.
The future is not just having an AI agent. The future is having an AI agent that is safe, useful, monitored, and aligned with your business.