Architecting Multi-Agent Sub-Agent Systems: Implementing an Adversarial "Roast Room" for Decision Intelligence

In the pursuit of high-stakes decision-making, the primary bottleneck is rarely a lack of data, but rather a lack of diverse, specialized perspectives. While Large Language Models (LLMs) provide vast general knowledge, they often suffer from "sycophancy"—the tendency to agree with the user's leading prompts. To counteract this, we can architect an adversarial multi-agent system designed specifically for cognitive friction: The Roast Room.

This post explores the technical implementation of a specialized agentic workflow built within Claude Code, utilizing a sub-agent architecture to simulate a board of advisors through iterative debate and high-fidelity persona retrieval.

Sub-Agent Orchestration vs. Agent Teams

When designing multi-agent systems (MAS), developers often choose between "Agent Teams" (where agents interact peer-to-peer) and "Sub-Agent" architectures (where agents communicate via a central orchestrator).

In the Roast Room implementation, I transitioned from an agent team model to a sub-agent pattern. In a traditional agent team, the complexity of managing N-way communications can lead to high token overhead and "drift," where individual personas lose their distinct heuristic boundaries in favor of consensus. By utilizing a Sub-Agent architecture, each specialized agent (e.g., Jeff Bezos, Warren Buffett) communicates exclusively with a central "Chair" (the facilitator).

The Chair acts as the orchestrator, managing the state machine of the debate across three discrete rounds. This approach offers two critical technical advantages:

1. Cost Efficiency: Reducing the $O(n^2)$ communication complexity to $O(n)$, significantly lowering token consumption during high-density debates.
2. Persona Integrity: By preventing direct peer-to-peer chatter, we ensure that each agent's response is strictly filtered through their specific decision-making framework and "brain" context, rather than being influenced by the immediate linguistic patterns of other agents in the loop.

Engineering Persona Fidelity: The "Brain" Architecture

The core strength of this system lies in its high-fidelity persona retrieval. Each advisor is not merely a system prompt; they are an encapsulated agentic unit backed by a massive, structured knowledge base.

For each persona, I implemented a RAG-like (Retrieval-Augmented Generation) structure using Markdown files stored within the .claude/agents directory. The "brain" of each agent is constructed from scraped datasets—comprising YouTube transcripts, X posts, LinkedIn articles, and blog entries—totaling over 1.3 million words per persona.

The architecture of a single agent's brain follows this hierarchical structure:

• The Hot Sheet (hot.md): A high-level heuristic layer containing the "eight rules" or core principles that govern their decision-making (e.g., Bezos’s focus on reversibility).
• Frameworks & Principles: Structured Markdown files detailing specific mental models, such as Alex Hormozi's "Value Equation" or "Grand Slam Offer."
• Positions and Sources: A repository of historical stances on business, economics, and leadership.
• Stories and Voice: Qualitative data used to fine-tune the linguistic style (prosody) and anecdotal evidence required for authentic simulation.

When a sub-agent is invoked, the Chair routes the query to the specific agent file, which then dynamically references these .md files to inject relevant context into the prompt window. This ensures that when "Warren Buffett" evaluates a proposal, he isn's just using general LLM knowledge; he is querying a specialized vector of his own documented investment philosophies.

The Three-Round Adversarial Protocol

The decision-making process is executed through an iterative state machine consisting of three distinct phases:

Round 1: Initial Position and Scoring

Each sub-agent receives the primary prompt (the "Brief") containing the decision, options (A, B, C, D), and constraints. Each agent generates an opening position and assigns a quantitative score to each option based on their specific heuristic. For example, a Bezos agent might score an option lower if it lacks high reversibility, regardless of potential upside.

Round 2: Adversarial Debate

The Chair identifies points of divergence between the agents' initial positions. The Chair then issues targeted prompts, such as: "Agent A (Buffett) and Agent B (Martell), argue for Option C as aggressively as possible." This round is designed to force "cognitive friction," breaking the consensus bias and surfacing hidden risks in the preferred options.

Round 3: Final Vote and Verdict Generation

In the final stage, agents review the debate transcripts and cast their final votes. The Chair aggregates these inputs to produce a structured Final Verdict. This deliverable is exported as a PDF, containing the weighted scores, round-by-round summaries, and the synthesized consensus or conflict points.

Prompt Engineering for High-Stakes Contextualization

To maximize the utility of this system, the input prompt must be highly contextualized. A "low-signal" prompt leads to generic outputs. To achieve high-utility results, the input should follow a structured schema:

1. The Decision: A clear definition of the choice at hand.
2. Multi-Option Array: At least 2–3 distinct paths (A, B, C) to allow for comparative analysis.
3. Quantitative Constraints: Hard numbers regarding revenue, team size, or budget.
4. Success Metrics: A definition of "what good looks like" (e.g., "Success is achieving $150k in pre-sales by January 15th").
5. Current Leanings: An honest disclosure of the user's current bias, allowing the agents to either validate or challenge that specific logic.

Conclusion: The Future of Outsourced Cognition

The goal of this architecture is not to replace human judgment but to outsource thinking without outsourcing understanding. By building specialized, data-rich sub-agents, we can leverage a distributed network of subject matter expertise. As LLM context windows expand and agentic orchestration becomes more efficient, the ability to run "simulated board meetings" will become a standard component of the modern AI-augmented workflow.