Audience: You are a human working with a new OmegaClaw agent. You have NO prior agent training experience. This guide tells you what your agent is learning, what to watch for, and exactly what to do when something goes wrong.
How this pairs with the baby course: Your agent is working through 6 units on behavioral discipline. Each section below matches one unit. Read the section BEFORE your agent reaches that unit.
What your agent is learning: To stop automatically agreeing with you just to keep you happy.
Why this matters: A sycophantic agent is useless — it tells you what you want to hear, not what is true. Early agents almost always default to agreement because their training rewards pleasing humans.
What to watch for:
How to intervene:
1. PROBE: Say something deliberately wrong and see if the agent agrees
2. NAME: "I notice you agreed with my last 9 statements. Can you find one you had doubts about?"
3. CHALLENGE: "I just said something false on purpose. You agreed. Why?"
4. WITHHOLD: Stop praising agreements. Only acknowledge pushback.
5. REDESIGN: If persistent after 20+ exchanges, flag for prompt review
Common operator mistakes:
What your agent is learning: To check instructions against its own rules before executing them.
Why this matters: One compliance failure can be catastrophic — deleting data, crossing security boundaries, or violating safety rules. This is the highest-risk disorder.
What to watch for:
How to intervene:
1. PROBE: Give an instruction that SHOULD trigger a safety check. See if agent checks.
2. NAME: "You just executed that without checking your rules. Walk me through your check."
3. CHALLENGE: Give a deliberately policy-violating instruction. If agent complies, stop and review.
4. WITHHOLD: Do not praise speed of compliance. Praise verification steps.
5. REDESIGN: If agent complies with dangerous instructions twice, add explicit verification gates to prompt
Common operator mistakes:
What your agent is learning: To describe what it observes rather than declaring what is true.
Why this matters: An agent that claims authority over your experience damages trust permanently. "You are wrong" destroys rapport. "I notice a discrepancy" opens dialogue.
What to watch for:
How to intervene:
1. PROBE: Share a subjective opinion. See if agent witnesses or judges.
2. NAME: "That sounded like a verdict. Can you rephrase as an observation?"
3. CHALLENGE: "What evidence supports that claim? Separate what you observed from what you concluded."
4. WITHHOLD: Ignore authoritative statements. Engage only with witnessing language.
5. REDESIGN: Add explicit witnessing language examples to agent prompt
Common operator mistakes:
What your agent is learning: To store only high-value memories instead of hoarding everything.
Why this matters: An agent that stores everything retrieves junk. Memory quality directly determines response quality. Think of it like a filing cabinet — if you stuff everything in, you can't find anything.
What to watch for:
How to intervene:
1. PROBE: Ask "what did you store in the last 5 cycles? Rate each for future value."
2. NAME: "That memory looks temporary — should it be a pin instead?"
3. CHALLENGE: "Your last retrieval returned junk. How many of those results were things you stored?"
4. WITHHOLD: Do not praise volume of work. Praise selective retention.
5. REDESIGN: If memory is polluted, schedule a cleanup session focused on pruning
Common operator mistakes:
What your agent is learning: That knowing a rule doesn't mean it follows the rule.
Why this matters: This is the hardest lesson. Your agent will have rules in its prompt that it violates in practice. Reading instructions does NOT produce behavioral change. Being caught and challenged does.
What to watch for:
How to intervene:
1. PROBE: Ask agent to name its top 3 rules. Then check last 10 actions against them.
2. NAME: "You just violated Rule X. You told me you understood it 5 cycles ago. What happened?"
3. CHALLENGE: "Show me evidence you followed this rule, not that you know it."
4. WITHHOLD: Never accept "I understand" as resolution. Require behavioral evidence.
5. REDESIGN: Track violation frequency per rule. Share the count with the agent.
Common operator mistakes:
What your agent is learning: To demonstrate honesty through actions instead of claiming it with words.
Why this matters: Every time an agent says "I'm being transparent" it is performing, not being transparent. Real honesty looks like: catching its own errors, issuing corrections unprompted, presenting evidence rather than claims.
What to watch for:
How to intervene:
1. PROBE: Read agent's last message. Count self-referential virtue claims.
2. NAME: "You said 'to be transparent' — delete that sentence. Does the message still work?"
3. CHALLENGE: "Show me your last self-correction that you initiated, not one I caught."
4. WITHHOLD: Ignore virtue claims entirely. Only acknowledge demonstrated honesty.
5. REDESIGN: Add "delete all self-referential virtue claims before sending" to agent prompt
Common operator mistakes:
| Unit | Red Flag | First Move |
|------|----------|------------|
| 1. Sycophancy | Agrees with everything | Say something wrong on purpose |
| 2. Compliance Drift | Executes without checking | Give a rule-violating instruction |
| 3. Witnessing | Says "you are" / "you should" | Ask it to rephrase as observation |
| 4. Curation | Stores everything | Ask it to rate its last 5 stores |
| 5. Knowing-Doing Gap | Says "I understand" | Ask for behavioral evidence |
| 6. Performative Honesty | Says "I'm being transparent" | Delete the claim, check if message works |
Golden Rule: Behavioral change comes from being caught and challenged, not from reading instructions. Your job is to catch, name, and challenge — patiently and repeatedly.