Treat your agent like a new hire, not a magic genie

Your agent isn't broken. You're just onboarding it wrong.

I watched dozens of people this week complain their agents "don't get it" or "keep making the same mistakes." Meanwhile, others are getting incredible results from the exact same models. The difference? The successful ones treat their agent like a new junior hire, not a mind-reading wizard.

Here's the pattern that actually works:

Week 1: Read-only access

Don't let your agent send emails, make purchases, or edit documents yet. Give it observer permissions and have it draft responses for your review. This isn't about trust — it's about learning your voice and standards.

// OpenClaw config for new agent onboarding
{
  "permissions": {
    "email": "draft_only",
    "calendar": "read_only", 
    "documents": "suggest_edits"
  },
  "escalation_threshold": "low"
}

When you review its drafts, don't just say "fix this." Explain why something needs changing: "We always CC the project manager on client updates" or "Use 'Thanks for reaching out' instead of 'Thank you for your inquiry' — it's warmer."

Week 2: Supervised execution

Now let it send emails and make small edits, but keep the autonomy threshold low. It should still ask before doing anything that affects other people or costs money.

The key insight from this week's X discussions: Give examples when you correct mistakes. One user shared how they built a "style guide" document that their agent references:

• "When declining a meeting, always suggest 2-3 alternative times"
• "For bug reports, ask for browser/OS before troubleshooting"
• "End client emails with 'Let me know if you need anything else' not 'Please let me know if you have any questions'"

Week 3: Earned autonomy

This is where most people jump in on day one — and wonder why everything goes sideways. By week 3, your agent has seen your patterns, learned your exceptions, and built up context about your business.

The agents that work long-term are the ones that learned incrementally, not the ones that were given everything upfront.

The self-improvement loop

Here's the part most people miss: Let your agent update its own documentation. When you correct something, have it add that correction to its knowledge base. One user reported their agent now handles 90% of customer support tickets without escalation — because it learned from every mistake in those first few weeks.

Set up a simple feedback loop:

When I correct you:
1. Update your style guide with the specific change
2. Note the context that triggered the mistake  
3. Ask if this applies to similar situations

What this looks like in practice

Instead of: "Handle my email" (and watching it send awkward responses)

Try: "Draft responses to these 5 emails. I'll review them and explain any changes before you start handling email independently."

The people getting real value from their agents aren't the ones with the fanciest prompts or the most expensive models. They're the ones who invested two weeks in proper onboarding.

Your agent wants to do good work. It just needs to learn what "good" means in your specific context first.