Seeing the 1.48 billion token consumption displayed in the terminal, I realized just how crazy the high-intensity debugging was this week. I initially thought setting up OpenClaw would mean it would run automatically, but the first two weeks were spent entirely babysitting—watching the AI repeat the same answer eight times or get stuck in endless debates about whether to use spaces or tabs.

The key to turning an AI agent from a toy into a productivity tool is to stop treating it like a chatbot and instead view it as infrastructure that needs careful design. Here are ten lessons learned from burning through 1.4 billion tokens.

### 1. Choosing the right model for daily tasks can save 80% of costs

Initially, I ran all tasks on Opus or Codex, even basic operations like heartbeat detection and status queries used the most expensive models. That was until I discovered the newly released Sonnet 4.6—it achieved 72.5% on the OSWorld benchmark, close to Opus 4.6's 72.7%, but at only one-fifth the cost.

If the budget is tight, you can switch to Kimi K2.5 on OpenRouter, costing about $0.6/$2 per million tokens. I recommend setting up a layered configuration: a primary model for daily tasks and fallback models for complex scenarios.

```json

{

"agents": {

"defaults": {

"model": {

"primary": "anthropic/claude-sonnet-4-6",

"fallbacks": [

"anthropic/claude-opus-4-6",

"openrouter/moonshotai/kimi-k2.5"

]

}

}

}

}

```

### 2. An AI without rule constraints will do all sorts of stupid things

A freshly installed OpenClaw will loop through failed solutions, mess up configuration files, and skip documentation to hallucinate solutions. The solution is to create a SKILL.md file in the workspace/skills/ directory to set behavioral guardrails for the AI.

The most effective rule is: **Must read documentation before making any changes**. Some netizens have developed the DocClaw skill, which forces the AI into a reconnaissance phase before executing code changes, directly halving the error rate.

### 3. Three files keep AI working continuously while you sleep

- **Todo.md**: A self-expanding task list. Give the AI a big task before bed, and it will break it down into subtasks and update the status.

- **ProgressLog.md**: A morning brief. Records the results of every build-test cycle and what was learned.

- OS configuration files: Define execution loops, planning discipline, and quality gate standards.

### 4. Overnight runs rely on cron jobs, not keeping windows open

Closing the session window causes the AI to lose all state. You need to set up cron tasks to wake the AI at specific times:

```bash

openclaw cron add --name "overnight-2am" --cron "0 2 * * *" --message "检查Todo.md，继续未完成任务"

openclaw cron add --name "overnight-4am" --cron "0 4 * * *" --message "更新进度日志"

openclaw cron add --name "overnight-6am" --cron "0 6 * * *" --message "总结通宵工作成果"

```

### 5. All important content must be persistently stored

Long sessions get compressed, and the AI quietly loses early decisions and state. The solution is to write all key information into markdown files in the workspace, like preparing onboarding documents for an employee who gets amnesia every morning.

### 6. Model capability is more important than the framework

Chat quality is completely different from agent quality. Some models can write poetry but choke when calling tools. Here is how current models perform in agent tasks:

| Model | Agent Quality | Tool Calling | Cost (per million tokens) |

|------|---------------|-------------|--------------------------|

| Claude Sonnet 4.6 | Excellent | Reliable | $3/$15 |

| Claude Opus 4.6 | Excellent | Reliable | $15/$75 |

| GPT-5.3-Codex | Excellent | Reliable | Pro Subscription |

| Kimi K2.5 | Good | Reliable | ~$0.6/2 |

| MiniMax M2.5 | Good | Reliable | $0.3/1.2 |

### 7. Start with one workflow, perfect it, then expand

Don't set up email + calendar + Telegram + web scraping + cron tasks all at once. Every integration is an independent point of failure. Start with a single morning brief cron job, and only add the next feature after it runs stably for a week.

### 8. Separate development from operations) for development: writing code, debugging, deploying features.

- OpenClaw for operations: monitoring, scheduling, communication, automation.

Keep the two separate to avoid context pollution.

### 9. A memory system prevents the AI from starting from scratch

OpenClaw has built-in vector memory capabilities, plus advanced solutions like Claw Vault and Supermemory. One developer's self-built Gigabrain system has indexed 911+ memories, allowing the AI to remember what methods work and what goes wrong, significantly boosting efficiency.

### 10. Security auditing must come first

OpenClaw has had real security incidents: a CVSS 8.8 remote code execution vulnerability, Bitsight and Censys teams found over 30,000 exposed instances, and a large-scale ClawHub supply chain poisoning attack (about 12% of the skill library was implanted with malicious code).

```bash

# 健康检查+自动修复

openclaw doctor --deep --fix --yes

# 安全审计

openclaw security audit --fix

```


One netizen commented: "It's only the fourth day, and I feel like a monkey with a gun." Many people have similar experiences when starting out. The key is to accept that the setup process *is* the work—writing Soul.md is product work, adjusting model routing is infrastructure work, and setting up cron tasks is operations work.

Behind the consumption of 1.4 billion tokens is the fact that AI agents can finally deliver real work results while you sleep. System setup is the moat; most people give up before reaching this stage.

As another user said: "This kind of transparency helps upgrade the entire ecosystem." Open-sourcing real experiences is more valuable than any perfect demo.