We just opened the waitlist for Something, and the part that surprised me most while building it wasn't the multi-agent orchestration — it was how hard it is to make an AI actually disagree.
Every model we tested defaults to being helpful, which in practice means agreeable. Even when explicitly prompted to "find flaws," the outputs would soften into "here are some considerations" instead of a real critique. We had to engineer around this specifically:
Separate system prompts with opposing reward framing — one agent optimizes for identifying growth potential, the other is explicitly told its only success metric is surfacing a disqualifying flaw
Structured output forcing a verdict, not a summary — the skeptic agent (Nothing) has to commit to a specific weakness category (unit economics, timing, technical feasibility) rather than hedging across all of them
A reconciliation step where both outputs get merged into one conviction score, so the founder isn't just reading two contradictory paragraphs
If anyone's built adversarial agent setups and hit the same "it just wants to agree with me" problem, curious how you solved it.
[Everyone who has a brain is a founder here]
Top comments (3)
You asked how others got past the "it just wants to agree" problem, so here's what tends to work for me: don't ask one model to critique, ask it to defend the opposite position it doesn't hold. "Argue that this startup fails" gets a real teardown where "find the flaws" gets hedging, because now agreeing is off the table, disagreeing is the assignment. Your forced-verdict-into-a-category trick is doing something similar, it removes the escape hatch of spreading soft doubt across everything. The one place I'd expect your setup to strain is that a skeptic told its only success metric is surfacing a disqualifying flaw will invent one when the idea is genuinely fine, which is its own kind of unhelpful. Does the reconciliation step have any way to let the skeptic come back with "honestly, nothing disqualifying here," or is it always forced to name a weakness?
That's why we aren't relying on the critic model only, in parallel there are two more models one to find good points and the other to judge that none of the models went overboard on the task so with time by using LLM evals the models will get to the right stopping point
This was a fascinating experiment. Building an AI agent that intentionally challenges ideas is a great way to uncover blind spots and strengthen reasoning rather than just seeking agreement. I also like that you shared both the technical lessons and the unexpected behaviors you encountered—it makes the post much more valuable for anyone experimenting with AI agents. Thanks for sharing your insights!