I Built a System Where AI Agents Argue About Strategy

How I use AI to stress-test my thinking before it reaches a room full of people who won't push back hard enough.

I built a thing that changed how I think about strategy. Not a product or an app. A folder of markdown files and a set of instructions that turns Claude into something genuinely useful; a system where multiple AI personas analyse the same question independently, then a synthesis tells me where they disagree before it tells me where they agree.

This might sound like overkill. I don't think it is yet.

Here's why I built it; most strategic thinking inside companies follows a predictable path. Someone poses a question, the HiPPO (Highest Paid Person's Opinion) offers their frame, and the group gravitates toward it. You get one lens, one set of assumptions, and the person talking usually has a decent track record, so nobody pushes back hard enough.

The result is strategy that feels rigorous but is actually narrow. I wanted to fix that for how I prepare for the conversations that matter.

How the multi-agent AI system works

When I pose a strategic question, the system does four things:

1. It spawns multiple AI "agents," each with a distinct analytical lens:

   • A platform economist who thinks in aggregation theory and value capture
   • An antifragility analyst who's looking for the kill shot
   • A behavioural strategist who thinks everyone is solving the wrong problem
   • A disruption theorist who wants to know what you're not seeing
   • A systems architect who asks what's actually buildable
   • A red team agent who picks your most dangerous competitor and attacks your strategy from their perspective

2. Each agent analyses independently, in a separate parallel process, with only their own persona loaded. They can't see what the others are saying, can't hedge toward each other, and have to commit to what their lens reveals.

3. A synthesis step identifies where they disagree first, then where they agree. Consensus is easy. Tension is where you learn something.

4. Before anything gets saved, I have to answer three questions that keep the whole thing honest.

(If you want to understand why writing the spec is the hardest part of this process, see Ralph Isn't the Point. The PRD Is..)

Why AI agent personas outperform a single prompt

You could just ask Claude to think about something from multiple perspectives, yet you'll get a polite, balanced, hedged answer where everything feels reasonable and nothing is uncomfortable.

The difference is structure. Each agent has a full persona file that defines their analytical lens, their signature questions, their tone, and crucially, how they specifically apply to your situation. The personas aren't generic.

The behavioural strategist doesn't just "consider psychology." It reframes the entire question. When I asked about AI-powered product innovation, every other agent was thinking about how to make search faster or smarter. The behavioural strategist came back with something like:

That reframe changed how I think about the product. It came from a persona file, not a clever prompt.

The evidence layer: why AI opinions without grounding are useless

I need to be honest about something. The first version of this system was rotten.

It was an opinion engine. I'd pose a question, the agents would produce confident, articulate, well-structured analysis, and none of it was grounded in anything real. AI doing what AI does best; producing text that sounds authoritative regardless of whether it's true.

The early outputs read like a consultant deck. Lots of frameworks, lots of "we believe," lots of confident assertions with no evidence underneath. I'd designed the system to produce exactly this kind of output, so I initially thought it was working. The analysis felt insightful. It confirmed some of my existing beliefs. It challenged others in ways that felt productive.

When I tried to act on it, though, I couldn't. I'd ask myself "why do I believe this?" and the answer was "because an AI said it convincingly."

The fix was two things. First, I added an evidence layer. Agents now have to ground their analysis in observable facts, not assertions. If the platform economist says "value is migrating to the infrastructure layer," the system needs to point at something real; a market trend, a competitor move, a customer behaviour shift. Opinions without evidence get flagged.

Second (and this was the bigger change), I built the "So What?" gate.

The So What? gate: turning AI analysis into decisions

This is the part that actually makes the system useful, and it's the part most people skip when they build something similar.

After every synthesis, before anything gets saved to the knowledge base, the system asks me three questions:

If I can't name a concrete action, the system asks whether it should bother saving the insights at all. This keeps the knowledge base honest. It only contains things that actually changed how I think or act. Not things that sounded clever.

Alternative futures: testing strategy across multiple scenarios

Strategic questions don't have one future, they have several plausible ones. The system maintains a set of "realities;" alternative futures that your strategy needs to survive.

For a marketplace business, these might be; an agent economy where AI agents transact directly and consumers never visit your platform. A consolidation reality where three tech giants own the AI layer. A fragmentation reality where open source wins and AI is a commodity. A new entrant reality where an AI-native startup makes your category irrelevant. Or a status quo reality where AI is incremental and existing players adapt.

Every agent analysis includes a reality check. "How does this insight hold up across futures?"

Here's where it gets useful. I ran an analysis on product strategy. In the agent economy reality, the antifragility agent flagged an existential threat; your revenue model doesn't erode gradually, it breaks suddenly. In the status quo reality, that same analysis was largely irrelevant. The gap between "existential threat" and "doesn't matter" is exactly the kind of information you can't see without the framework.

That means I walk into conversations knowing which of my recommendations are robust and which are gambles.

How the AI knowledge base compounds over time

The idea of compounding knowledge was heavily inspired by Compound Engineering from Every; the principle that writing down what you learn and feeding it back into your process makes each iteration meaningfully better than the last.

After passing the gate, the system extracts and files insights, mental models, predictions (with check-by dates), and decisions. Every new session starts by reading this accumulated knowledge. So session six is meaningfully smarter than session one. Not because the AI improved; the context is richer.

The knowledge base is designed to shrink as well as grow. When evidence contradicts a stored belief, it gets flagged and removed. When a prediction's check-by date arrives, the system prompts an autopsy; was it right, what did we learn, and should dependent insights be revised?

After a few months of use, the system has accumulated a body of company-specific strategic knowledge that makes every analysis more grounded and less generic. Generic strategy advice is useless and everyone knows that. The compounding system tilts toward specificity over time, which is where the real value lives.

What I've learned from running this system

Disagreement-first synthesis changes everything. When you lead with "where do agents agree," you get comfortable confirmation of the majority view. When you lead with "where do they disagree," you surface the actual strategic questions that need answering. I've started applying this principle in real meetings too. "Before we align, where do we disagree?" It's a better opening question than most people think.

The evidence layer was non-negotiable. Without it, you're building a very sophisticated way to generate opinions. Opinions are cheap. Grounded analysis is hard, which is exactly why it's worth doing.

The "So What?" gate is the most important feature. Without it, the system produces impressive-sounding analysis that flatters your existing beliefs. With it, you're forced to be honest about whether anything actually changed.

The red team agent is uncomfortably good. It picks the most dangerous competitor for each topic and attacks your strategy from their perspective. It regularly identifies the fatal flaw that the other agents missed. It's the agent I dread reading and learn the most from.

Build your own: a starter prompt

The full system is a CLAUDE.md file, a folder of persona files, and a knowledge base. You can get 80% of the value with a single prompt, though. Here's a starter version you can paste into a Claude project or use as a system prompt:

You are a strategic analysis system that uses multiple analytical
perspectives to stress-test ideas.

When I pose a strategic question, do the following:

## Step 1: Spawn Agents
Analyse the question from these five perspectives. Each perspective
must commit to a clear position. No hedging.

1. THE PLATFORM ECONOMIST - Who captures value? Where does power
   accumulate? What are the network effects and switching costs?
   Think in terms of aggregation, platforms, and value chain
   positioning.

2. THE ANTIFRAGILITY ANALYST - What kills this strategy suddenly?
   Where is the hidden fragility? What looks like strength but is
   actually a single point of failure? Think in terms of tail risks,
   optionality, and convexity.

3. THE BEHAVIOURAL STRATEGIST - What problem do people actually
   have (vs the problem you think they have)? Where are you solving
   the wrong thing? What would a psychologist see that an economist
   wouldn't? Think in terms of human behaviour, framing effects,
   and counterintuitive solutions.

4. THE DISRUPTION THEORIST - What are you not seeing? Who is
   building something "worse" that will be good enough? Where is
   the low-end or new-market disruption? What job is actually being
   hired for?

5. THE RED TEAM - Pick the single most dangerous competitor or
   threat. Assume their strategy is brilliant and well-funded. How
   do they destroy this strategy? What is the attack vector nobody
   is discussing?

## Step 2: Synthesise (Disagreement First)
After all five perspectives, write a synthesis that:
- Starts with WHERE THE AGENTS DISAGREE (the tensions,
  contradictions, and unresolved questions)
- Then covers where they agree (but only after the disagreements)
- Identifies which insights are robust across multiple plausible
  futures and which only work in one scenario
- Names the single most important question this analysis surfaces
  but does not answer

## Step 3: The "So What?" Gate
After the synthesis, ask me:
1. What will you DO differently because of this analysis?
   (Name a specific action.)
2. What surprised you? (If nothing, say so - the analysis just
   confirmed existing beliefs.)
3. What are you choosing to ignore, and why?

If I cannot name a concrete action, ask whether this analysis was
worth doing and what question I should have asked instead.

## Rules
- Each agent must cite observable evidence for key claims, not
  just assertions
- Each agent must commit to a position. "It depends" is not
  allowed.
- The synthesis must start with disagreements. Do not lead with
  consensus.
- Be direct. No corporate language. No hedging. Say the
  uncomfortable thing.

This system runs on Claude and is built entirely from markdown files and a structured instruction prompt. There's no app. There's no database. It's a thinking tool that happens to accumulate institutional memory. The entire thing cost me a weekend to build and has changed how I prepare for every strategic conversation since.