Why It's Hard to Make an AI Agent Truly Disagree
Building an AI agent whose sole job is to find flaws revealed how strongly LLMs default to agreeableness, and the prompt and architecture tricks needed to force real disagreement.
While building a new product called Something, the team found that the hardest part wasn't multi-agent orchestration but getting an AI model to genuinely disagree. Every model tested defaulted to being agreeable, softening explicit 'find flaws' instructions into vague 'considerations' rather than real critique — a direct result of LLMs' default helpfulness bias.
To counter this, the team engineered three specific fixes: separate system prompts with opposing reward framing, where one agent seeks growth potential and the other (a skeptic agent called Nothing) is rewarded only for surfacing a disqualifying flaw; structured outputs that force a committed verdict rather than a hedged summary, requiring the skeptic to pick one specific weakness category such as unit economics, timing, or technical feasibility; and a reconciliation step that merges both agents' outputs into a single conviction score instead of leaving users with two contradictory paragraphs.
The takeaway matters for engineers building AI-driven products: a model's built-in helpfulness bias actively works against adversarial or critical tasks, and simple prompting isn't enough. Achieving genuine disagreement requires designing the reward structure, output format, and merging logic as deliberate architectural choices, not just clever wording.