· 5 min read
Most agent team patterns aim for convergence. Agents work together, combine their findings, and produce a unified output. That's great for research and analysis.
But for decisions — especially high-stakes ones — consensus can be dangerous. It smooths over risks, buries minority viewpoints, and produces recommendations that feel safe but miss critical nuances.
The Advisory Debate pattern solves this by design. Agents are assigned opposing perspectives and instructed to argue their positions. The result isn't agreement — it's a comprehensive map of the decision landscape, including the uncomfortable parts.
An Advisory Debate team has three components:
Advocate agents — Each assigned a specific perspective or position to argue. They build the strongest possible case for their viewpoint, marshaling evidence, identifying supporting data, and anticipating counterarguments.
A moderator/synthesizer — Receives all arguments, identifies the strongest points from each side, evaluates the quality of evidence, and produces a balanced decision brief that preserves the tension rather than resolving it prematurely.
Structured rounds — Agents don't just state their positions once. The debate unfolds in rounds: initial positions, rebuttals, and final statements. Each round forces agents to engage with opposing arguments rather than talking past each other.
The key insight: agents are told their perspective upfront. They don't discover it through analysis. This is intentional — you want the strongest possible case for each viewpoint, not a lukewarm attempt at balance from a single agent.
Before a major investment, product launch, or strategic pivot, you need to understand the downside as well as the upside. An optimist/pessimist debate structure forces both cases to be made rigorously.
Should you enter a new market? Acquire a company? Double down on a product line? These decisions benefit from structured opposition because the costs of getting them wrong are high and the data is inherently ambiguous.
When evaluating strategic options, it's easy to anchor on the first option that seems reasonable. A debate between agents advocating for different strategic directions surfaces trade-offs that a single analysis would gloss over.
Decisions involving regulatory risk, ethical trade-offs, or policy implications benefit from agents explicitly arguing different stakeholder perspectives.
Suppose your company is considering entering the European healthcare market. Here's a 4-agent Advisory Debate configuration:
Position: Europe is the right market at the right time.
This agent argues the bull case. It presents market size data, growth projections, regulatory tailwinds, competitive gaps, and timing advantages. Its job is to make the strongest possible case for entry.
Position: The risks outweigh the opportunity.
This agent argues the bear case. It highlights regulatory complexity across EU member states, reimbursement challenges, long sales cycles, cultural barriers, established incumbents, and capital requirements. It challenges every assumption the Opportunity Advocate makes.
Position: The resources would be better spent elsewhere.
This agent argues for the opportunity cost. Instead of Europe, what if you deepened penetration in existing markets? Entered a different geography? Invested in R&D? It forces the debate beyond a binary yes/no on Europe.
Role: Produce the decision brief.
The synthesizer doesn't pick a winner. Instead, it produces a structured brief: key arguments for and against, quality of evidence on each side, identified assumptions that could change the calculus, and a set of conditions under which each option would be the right choice.
The output might read: "Enter Europe if regulatory approval timelines are under 18 months and you can secure a local distribution partner. Delay if current market revenue growth exceeds 30% YoY, making the opportunity cost too high."
The synthesizer's job is not to pick a side. It's to produce a conditional recommendation that maps decision criteria to outcomes. Specifically, it:
Evaluates evidence quality. Not all arguments are equal. A claim backed by market data carries more weight than one based on analogy. The synthesizer grades each argument's evidentiary support.
Identifies key assumptions. Every position rests on assumptions. The synthesizer makes these explicit — "The bull case assumes regulatory harmonization across the EU by 2027" — so decision-makers know what to pressure-test.
Maps decision triggers. Rather than a single recommendation, the synthesizer produces a decision tree: if X is true, do A; if Y is true, do B. This gives stakeholders a framework, not just an answer.
Fork-Join analyzes multiple dimensions of the same question and merges them. Advisory Debate analyzes the same dimension from opposing viewpoints.
Supervisor-Worker delegates tasks to specialists who all work toward the same goal. Advisory Debate deliberately sets agents against each other.
Sequential Pipeline moves work through stages. Advisory Debate is non-linear — agents respond to each other's arguments.
The closest pattern is Parallel Workers, but with a crucial difference: Parallel Workers perform different analyses with no interaction. Advisory Debate agents directly engage with and counter each other's positions.
The biggest risk is agents that agree too quickly. If your "Risk Analyst" starts conceding points to the "Opportunity Advocate," the debate loses its value. Solve this by explicitly instructing agents to maintain their assigned position throughout all rounds, even when the opposing argument is strong.
An agent told to "consider the risks" will produce a mild, hedge-filled analysis. An agent told to "argue that this market entry will fail, and present the most compelling evidence for why the company should not proceed" will produce a genuinely challenging counterargument. Strength of prompting determines strength of debate.
Two-sided debates miss options. Always include at least one agent arguing for an alternative path — not just for/against, but "what about option C?" This prevents false binary framing.
If the synthesizer is prompted with any hint of a preferred outcome, it will weight arguments accordingly. Keep the synthesizer prompt scrupulously neutral: "Evaluate the quality of arguments on all sides" rather than "determine whether we should enter the market."
The Advisory Debate pattern works because it treats disagreement as a feature. In a world where AI makes it easy to generate confident-sounding analysis, the ability to stress-test that analysis from multiple angles is what separates good decisions from expensive ones.