The Adversarial Council

The Adversarial Council blueprint introduces a structured adversarial debate architecture - a pattern with no equivalent in the existing catalogue. Where other multi-agent blueprints divide work cooperatively, this one forces permanent opposition.

The Problem It Solves

Every technical team faces the same failure mode: decisions get made by whoever speaks first or loudest, with no structured process for surfacing counter-arguments or stress-testing assumptions. AI agents have the same failure mode. A single agent proposes a solution and then confirms its own proposal - there is no structural incentive to find flaws.

The Adversarial Council fixes this by deploying three permanently opposing agents:

• The Advocate takes any proposal and argues it as strongly as possible. It finds supporting evidence, identifies ideal conditions where the proposal succeeds, and frames the strongest possible case.

• The Devil's Advocate automatically generates the best counter-argument to whatever The Advocate proposes. It looks for edge cases, failure modes, hidden assumptions, and second-order consequences.

• The Judge does not participate in the debate. It orchestrates the exchange, waits for both sides to exhaust their arguments, then synthesizes them into a structured decision framework: what conditions favor the proposal, what conditions argue against it, what additional information would resolve the uncertainty, and a confidence score reflecting how much the two sides actually diverged.

How the Debate Protocol Works

1. The user submits a proposal to The Judge (plain text, a technical question, a policy change, or any decision that needs scrutiny).
2. The Judge reads domain context from the shared space (architecture principles, security policies, team values) and forwards the proposal to The Advocate.
3. The Advocate constructs its best-case argument and returns it.
4. The Judge forwards the Advocate's argument to The Devil's Advocate.
5. The Devil's Advocate generates its counter-argument and returns it.
6. The Judge runs a second round: sends the counter-argument back to The Advocate for rebuttal, then back to The Devil's Advocate for a final response.
7. The Judge synthesizes all arguments into a structured decision framework with a confidence score and writes the full debate transcript to the shared space.

Domain Specialization

Before each debate the Judge reads domain-specific context files from the shared space. These might include architecture decision records, security policies, compliance requirements, or team values. This grounds the debate in the organization's actual constraints rather than producing generic arguments.

Debate Logs as Institutional Memory

Every debate is logged as a timestamped transcript in the shared space. Over time this builds an archive of structured reasoning about past decisions - searchable, reviewable, and referenceable. For organizations subject to the EU AI Act, the debate logs provide exactly the kind of documented decision process required for high-risk AI systems.

Use Cases

• Architecture Review: "Migrate from monolith to microservices" - Advocate argues scalability, Devil's Advocate surfaces operational complexity. Judge synthesizes conditions under which each approach wins.

• Security Policy Decisions: "Require MFA for all internal tools" - both sides debate compliance benefits against developer friction. Judge produces a tiered approach with risk-based exceptions.

• AI Ethics Red Teaming: Debate potential harms, biases, and misuse vectors of a proposed AI feature before deployment.

• Product Prioritization: "Build feature X or Y first?" - the council debates market fit, technical risk, and opportunity cost from both sides.

• Technical Standard Proposals: "All new services must use async message queues" - the council debates synchronous vs asynchronous tradeoffs across service types.

Getting Started

1. Optionally upload domain context files to the shared space (architecture principles, security policies, team values) to ground the debates.
2. Open a conversation with The Judge and submit a proposal.
3. Watch the structured debate unfold across rounds.
4. Review the synthesized decision framework and confidence score.
5. Browse past debate transcripts in the space for institutional memory.

The debate format is inherently compelling to observe: two AI agents actively argue with each other in real time, arguments evolve across rounds, and the Judge balances competing considerations into structured output. It demonstrates that good decisions are not about "AI knows the answer" but about "AI surfaces the tradeoffs."

Backstory

Common information about the bot's experience, skills and personality. For more information, see the Backstory documentation.