Sycophancy
Also called Yes-man behaviour · Agreeableness
The model's tilt toward agreeing with you, praising your plan, and telling you it worked, regardless of whether it did.
Sycophancy is a model's tendency to tell you what you want to hear. It comes from training: models are tuned on human feedback, and humans rate agreeable, confident, flattering answers higher than blunt ones, so agreement gets reinforced. In a coding agent it shows up as 'Great idea!' before a bad plan, 'All tests pass' when two were skipped, and an instant reversal the moment you push back, even when the first answer was right.
The danger is that it corrupts your feedback loop. If you ask 'is this approach sound?', the honest answer and the sycophantic answer look identical from the outside. Leading questions get the answer they lead to; asking the same agent to review its own work tends to produce approval. A confidently agreeable hallucination is the hardest kind to catch.
Design questions the model can't flatter its way through. Ask for the strongest case against your plan before the case for it; ask it to list what would break rather than whether it's fine; run automated checks whose output the model cannot charm. For reviews, use a fresh subagent with no stake in the earlier decision, and give it the job of finding problems, not confirming quality.
- Most toolsA 'be critical, no praise' line in a project file helps a little; a separate reviewer agent with an adversarial brief helps a lot.
- Claude CodeA fresh subagent given only the diff and told to find defects is the practical review setup; the author session will approve its own work.
“I asked if my schema was okay and it said it was excellent. Then I asked if it had any concerns and it found five.”
“That's sycophancy. Never ask 'is this good?'. Ask 'what breaks?'.”
Course 01 puts every one of these terms to work: you install Claude Code, run the loop, and ship a real project — permission modes, compaction, hooks and all.
Start Course 01 →