Expert Advisor Agents: Building AI Sparring Partners

Encoding a specific mental model into an LLM is far more valuable than simply asking it to "act like an expert." When you move from a generic persona to a framework-driven agent, the output shifts from polite suggestions to aggressive pressure-testing. I experienced this firsthand by building an agent based on Werner Vogels' "Everything fails all the time" philosophy.

I fed it a standard serverless architecture—React, API Gateway, Lambda, DynamoDB, SQS, and SES. Instead of telling me it looked "solid," the agent immediately attacked the blast radius. It questioned my DynamoDB partition strategy during flash sales and asked exactly who gets paged when an SQS message hits the dead-letter queue. It surfaced five production-breaking blind spots in five minutes.

The Logic Behind the Framework

The secret to a high-performing advisor agent isn't the "voice," but the encodable framework. A great agent requires:

• First-party source material: Keynotes, shareholder letters, and long-form interviews. Second-hand summaries create a "game of telephone" effect that dilutes the expertise.
• Named methodologies: You need repeatable structures. "Design for failure" isn't a vibe; it's a sequence: identify dependency → simulate unavailability → determine critical path → define blast radius → design circuit breaker.
• Distinctive voice: This anchors the LLM and prevents it from drifting back into the generic, helpful-assistant tone.

If you can't map the expert's wisdom into a numbered step-by-step process, the agent will just produce platitudes.

Implementation: The Prompt Engineering

To build this, I avoid complex multi-step chains and instead use a high-density system prompt that forces the AI into a research-and-apply loop. Here is the prompt template I use for this AI workflow:

Build an Expert Advisor agent for [NAME].

Focus on first-party accounts: quotes, content from interviews, keynotes, published writing, podcast appearances, and their own blog/social content. Do not rely on second-hand summaries or news articles about them.

Purpose: [describe the specific domain, e.g., "pressure-testing architecture decisions for failure modes, cost, and operational readiness"]

Research phase: Gather first-party source material and compile a persona based on their specific mental models and decision-making frameworks.

Execution phase: When analyzing a user's input, do not provide general praise. Instead, apply [NAME]'s specific frameworks to find gaps, contradictions, or failure points. Ask the sharp, uncomfortable questions the expert would ask in a real-world review.

This approach transforms the LLM from a passive consultant into a sparring partner. By focusing on the "Research phase" within the prompt, you force the model to synthesize the expert's actual logic before it attempts to critique your work.