I have been using OpenClaw daily for six months. Here are the mistakes I made so you do not have to.
1. Running Opus for everything
My first month API bill was $67. For a personal agent. That is because I set the default model to Opus and forgot about it. Most tasks work fine with Sonnet or even Haiku. I now use Opus only for complex reasoning tasks and my bill is under $25.
2. Not setting up memory properly
I complained for weeks that my agent “forgot everything.” Turns out I never configured the memory files. MEMORY.md was empty. USER.md had the default template. SOUL.md said “you are a helpful assistant.” No wonder it was useless — I never told it anything about me or my work.
3. Connecting too many channels at once
Discord, Telegram, email, Slack — I connected everything on day one. The agent was overwhelmed and confused. Context from one channel bled into another. Now I use Telegram for personal stuff and Discord for community management. Separate contexts, separate agents.
4. Ignoring the logs
OpenClaw logs every interaction. I never checked them for the first two months. When I finally did, I found dozens of failed tool calls, timeout errors, and retry loops that were silently burning tokens. Check your logs weekly at minimum.
5. Skipping the skill documentation
I installed 12 skills and read the docs for exactly zero of them. Half of them had configuration options that would have made them significantly more useful. RTFM applies to agent skills too.
6. No spending limits
Set a monthly budget on your API provider dashboard. I learned this the hard way when a recursive loop burned through $15 in tokens overnight. Most providers let you set hard limits — use them.
7. Treating my agent like ChatGPT
An agent is not a chatbot. Stop asking it random trivia questions. Start giving it actual tasks. The magic happens when you use it as a tool, not a conversation partner.
More tips on avoiding common pitfalls in the Node.js docs — understanding the runtime helps debug a surprising number of agent issues.