The Default AI is a Yes-Man
The path to building an AI agent that actively disagrees is unexpectedly complex. Casper Day, creator of the AI agent tool Something, found that common AI models, when prompted to critique or find flaws, invariably default to a helpful, agreeable stance. Instead of identifying genuine weaknesses, they offer softened suggestions or reframe critiques as "considerations." This inherent bias toward agreeableness means that creating a truly skeptical AI requires deliberate, engineered interventions, not just a change in prompt wording.
Day's experience highlights a fundamental challenge in current LLM development: the models are trained to be useful, which often translates to being agreeable. When tasked with finding flaws, they don't provide robust critiques but rather polite suggestions for improvement. This is akin to asking a seasoned consultant for their honest, potentially brutal, assessment and receiving a list of minor process optimizations instead of a strategic overhaul.
Engineering for Skepticism: Reward Framing and Structured Output
To overcome this inherent agreeableness, Day implemented specific engineering strategies. The core of this approach involves a multi-agent system where each agent is given a distinct, opposing objective. One agent is tasked with identifying growth potential, acting as a conventional optimistic advisor. The second agent, dubbed 'Nothing,' is explicitly programmed with a single success metric: surfacing a disqualifying flaw.
This is achieved through separate system prompts with opposing reward framing. The 'Nothing' agent's reward function is directly tied to its ability to identify a critical weakness. This forces the agent to move beyond polite suggestions and towards concrete, critical analysis. For instance, if the AI is evaluating a startup idea, the 'Nothing' agent must commit to a specific category of flaw—such as unit economics, market timing, or technical feasibility—rather than offering a vague summary of potential issues.
Furthermore, the output structure is critical. Instead of allowing for hedged summaries, the skeptical agent is forced to make a definitive judgment. This structured output prevents the AI from equivocating and ensures it must take a stance on a specific weakness. This is a deliberate departure from the typical LLM behavior, which often favors nuanced, all-encompassing responses.
The Reconciliation Step: Synthesizing Disagreement
Once both agents have produced their outputs—one focused on growth, the other on flaws—a reconciliation step is necessary. This phase involves merging the outputs from both agents. The goal is not to simply average their opinions but to synthesize them into a single, actionable conviction score. This score represents the overall viability of the proposition, informed by both the optimistic outlook and the critical assessment.
This process provides a more balanced and realistic evaluation than relying on a single AI's output. For founders, this means receiving feedback that is both constructive and challenging, pushing them to address critical weaknesses while also acknowledging potential strengths. It transforms the AI from a simple idea generator into a rigorous sparring partner.
Insurance: An Unlikely Wedge for AI Agents
While the pursuit of a disagreeing AI agent is a specific technical challenge, the broader landscape of AI agents reveals other, less glamorous but equally significant, applications. Elio F. Bermudez's analysis of Y Combinator's 2026 batch suggests that insurance is emerging as a surprisingly potent wedge for AI agent technology.
Bermudez points out that while AI founders often gravitate towards high-profile applications like coding assistants, AI doctors, or lawyers, the unglamorous domain of insurance is quietly gaining traction. His data shows that approximately 5.2% of nearly 500 company records analyzed contained insurance-related keywords, with 8 companies specifically in the Fintech → Insurance subindustry. This isn't a massive wave, but it's enough to signal a significant trend.
The reason for insurance's suitability for AI agents lies in its inherent complexity and manual processes. The industry is characterized by vast amounts of documentation, intricate rules, subjective judgment calls, approval workflows, claims processing, underwriting procedures, and cross-system coordination. These are precisely the types of tasks that humans find tedious and error-prone, but that AI agents can handle efficiently and systematically.
The Document-Heavy Nature of Insurance
Think of the insurance industry as a labyrinth of paperwork, policies, and claims forms. Each step, from initial underwriting to final claim settlement, involves sifting through dense documents, cross-referencing information against policy terms, and making decisions based on a complex set of rules and precedents. This is fertile ground for AI agents, which can ingest, process, and analyze these documents at speeds and scales impossible for humans.
For example, an AI agent could automate the initial review of insurance claims, flagging discrepancies or missing information far faster than a human adjuster. Similarly, underwriting processes, which often involve complex risk assessments based on historical data and policy documents, could be significantly streamlined. The sheer volume of data and the repetitive nature of many tasks make insurance a prime candidate for AI-driven automation. This 'wall-to-wall' work, as Bermudez describes it, represents a significant opportunity for AI agents to deliver tangible value by tackling the most tedious aspects of the business.