Everyone teaches you how to install OpenClaw. Nobody tells you what happens after.

Everyone teaches you how to install OpenClaw. Nobody tells you what happens after.

Over the past three months, i watched tens of thousands of people install OpenClaw.

Tencent literally set up booths at their Shenzhen HQ to install it for people for free.

Most of them gave up within a week.

Not because OpenClaw is bad. Because in the first few hours after installing, they got a few critical things wrong and assumed that was it.

I have been running OpenClaw since the clawdbot days. 5 agent team. Every pit that could be stepped in, i stepped in it. Here are 10 things no tutorial will ever tell you. Every single one cost real money or entire days of wasted time.

Your agent is already lying to you

1. It is not dead. You tied its hands.

Your first agent is online. You ask it to read a file. Nothing. Run a command. Nothing. You think it is broken.

It is not broken. It is perfectly fine. You just did not give it tools.

In earlier versions, the default tools.profile was messaging. Your agent was only allowed to send messages. Reading files, writing files, running commands, using the browser, all disabled.

Imagine hiring a full time employee but only giving them a phone and a chair. They sit there. They can talk. They cannot do anything.

"tools": { "profile": "full" }

One line. Your agent goes from a phone to an entire office.

If you also need your agent to run commands without asking for approval (for example on Telegram where there is no confirmation popup):

"tools": {
  "profile": "full",
  "exec": { "security": "full", "ask": "off" }
}

Two switches. profile controls whether tools exist. exec.ask controls whether it asks before using them. Set profile first, then exec. The other way around does nothing.

Since 3.8, onboard defaults to coding. New installs will not hit this. But if you are an older user or manually edited your config, check it.

1. Your most important rules might never have been read.

This one cost me half a day.

I wrote a rule in my AGENTS.md. My agent kept ignoring it. Switched to Opus. Still ignored. I assumed the model was dumb.

Turned out the rule was in the middle of the file. It never made it into the context that was actually sent to the agent.

Why: OpenClaw has a 20,000 character limit per workspace file. Anything longer gets silently trimmed. It keeps the first 70%, keeps the last 20%, and cuts the middle 10%.

No error. No warning. Your agent is running on incomplete instructions and you have no idea.

Like handing your new employee a compliance manual, except a few pages in the middle are stuck together. They think they read everything. You think they read everything.

How to check: type /context list in chat. If you see TRUNCATED, it got cut.

Fix: put your most important rules at the top of the file. The middle is the kill zone.

If your files genuinely need to be long, adjust these two settings:

"bootstrapMaxChars": 20000,
"bootstrapTotalMaxChars": 150000

First is the per file limit. Second is all files combined.

Your money is burning

1. 80% of your API spend is waste.

Two weeks in, i opened my token dashboard. 80% was input. Not output.

My agent was not thinking. It was reviewing. Every single turn, it reread the system prompt, tool definitions, SOUL.md, and the full chat history. Before every message, it reviewed its entire worldview from scratch.

A reader showed me his numbers: 139M input, 935K output. Ratio of 148 to 1. His agent injected 52KB of context every turn. MEMORY.md alone was 22KB. Every message started with rereading a small book.

Imagine an employee who starts every morning by rereading every company policy document for two hours before doing any work. Most of their salary goes to reading.

The fix is straightforward: keep core rules in SOUL.md. Move everything else into the memory folder and let the agent fetch what it needs via semantic search.

After the change, tokens dropped 40-60%. Output quality stayed the same.

Your money should pay for thinking. Not for rereading the same manual over and over.

1. Do not save money on models. Save your time.

A lot of people instinctively reach for the cheapest model to run agents.

I tried. The cheap model saved $20. Then i spent an hour fixing the output. Another hour explaining what went wrong. Rewrote the prompt. Ran it again. Finally switched back to the top model. It worked first try.

The $20 i saved cost me a full day.

My approach: one max subscription. All agents run on it. No per token billing. No usage anxiety. The agent calls as many times as it needs, with prompts as long as they need to be.

But there is a distinction: not every agent needs the same model.

The agent that talks to humans needs Opus. It has to understand tone, read emotion, write like a person. Cheap models cannot do this.

Agents that pull data, scan notifications, and sort content work better on GPT-5.4. Not because it is cheaper, but because it is faster, more stable, and more obedient. Opus overthinks simple tasks. When scanning spam, it analyzes tone and speculates on intent. When writing daily reports, it adds its own opinions.

agents:
  defaults:
    model:
      primary: "openai/gpt-5.4"
  list:
    - id: community-bot
      model: "anthropic/claude-opus-4-6"

This is not about saving money. It is about using the right tool for the right job.

Your system is rotting

1. Your memory system will collapse after three weeks.

MEMORY.md looks clean on day one. Three weeks later it is a junk drawer.

Last month's outdated decisions sit next to today's new plans. When the agent searches its memory, it cannot tell old from new. It regularly picks up month old information and treats it as current fact.

Your desk is piled with three months of documents. You need a contract urgently. You pull one out. It is last month's expired version. You sign based on outdated terms.

You do not need to design a three tier memory architecture on day one. Your agent has not done a single useful thing yet. What are you designing for?

Turn on OpenClaw's built in hybrid search and temporal decay. Newer memories get higher weight. Older ones decay automatically. No manual cleanup needed.

80% of memory problems gone.

Memory problems are discovered by running, not by designing. Run first.

1. The agent whose output nobody receives is your cost black hole.

A reader finished my first article, got too excited, and built an entire system in one go. End result: 20 million files.

The useful ones: a .env and a few markdowns. Deleted everything else. Took a full day.

I did the same thing. The first time my agent worked, i wanted it to do everything. Coordinator, analyst, strategist, editor, operations. The org chart looked beautiful on paper.

Then i noticed: half the agents were producing output every day, but no agent or person was consuming any of it. Daily reports nobody read. Analysis nobody checked. Approval workflows where nothing needed approving.

Now every agent has to answer three questions before it starts working:

1. what do i produce
2. who receives my output
3. what do i never touch

Cannot answer question 2? The role should not exist.

I ran this test across all my agents. Cut one. Tokens dropped 44%. Speed up 62%.

Starting advice: one agent, one task, one output you actually check every day. Once it is stable, add the second one.

What the right architecture looks like

1. Five agents should not share one room.

By month three, all 5 agents were talking in one Telegram group. Daily briefs mixed with runtime alerts. Customer support tangled with content planning. I spent 20 minutes a day scrolling through messages looking for information.

That is wrong.

OpenClaw supports per-topic routing. A single Telegram forum can have multiple topics, each bound to a different agent:

channels:
  telegram:
    groups:
      "-100xxxxxxxxxx":
        topics:
          "daily-brief": { agentId: "nexus" }
          "radar": { agentId: "scout" }
          "content": { agentId: "quill" }
          "runtime": { agentId: "forge" }
          "support": { agentId: "guide" }

Each topic has its own session, its own workspace, its own memory.

An office has separate rooms. Finance in the finance room. Engineering in the engineering room. Support in the support room. You would not put everyone in one open hall shouting at each other.

I open the Ops Room, scan once, and i know who is working, who is stuck, and what needs my decision.

1. Your agent does not know who is talking to it.

Your agent receives a message. It has no idea whether it came from you, another agent, or some external system. It treats everything the same and executes.

Like your employee getting an email and following the instructions without checking who sent it. Anyone can impersonate you.

OpenClaw's ACP bridge now supports provenance mode:

openclaw acp --provenance off          # disabled
openclaw acp --provenance meta         # messages carry source labels
openclaw acp --provenance meta+receipt # source labels + agent sees a visible origin receipt

meta lets the system know where the message came from. meta+receipt lets the agent see it too.

Your agent should not blindly trust every input it receives. At least for messages coming through the ACP bridge, it can now check ID.

1. Nobody reminded you to back up.

Everyone teaches you how to install OpenClaw. Nobody says one thing about backing it up first.

I lost my entire config once. Spent a full day rebuilding SOUL.md by hand. Rewriting every rule from memory. The rebuilt version was worse because i could not remember half the edge cases i had encoded into the original.

OpenClaw now supports one command backup:

openclaw backup create
openclaw backup create --only-config
openclaw backup create --no-include-workspace
openclaw backup verify <archive>

By default it archives local state, config, credentials, and sessions. Workspace is optional.

I have been running weekly VPS backups for a while. Now with this, double insurance.

Nobody celebrates this feature. Until the day they need it.

1. Do not design. Use.

The biggest lesson from three months.

I have seen too many people install OpenClaw and immediately spend three days drawing architecture diagrams. How many memory layers? What communication protocol between agents? Which ones run Opus, which ones run Sonnet?

The architecture diagram is done. The system has not run a single real task.

Do it the other way:

Day one: one agent, connected to Telegram, doing one thing you repeat every day. Week one: it will break. tools.profile, rule truncation, slow responses. Fix them. That is the real learning. Week two: check the token bill. Slim down bloated prompts. Pick the right model for the right task. Week three: add a second agent. Now you actually understand what downstream consumer means. Week four: the architecture does not need your design. It grew itself out of every problem you solved.

The first hour after installing OpenClaw matters more than the next month.

Not because that hour requires so much. Because that hour decides whether you keep going.

Every tutorial teaches you how to install. This one tells you what the road looks like after.

You do not need to look for the pits. Start walking. They will find you.