Skip to content

ksmeltzer/ARC-7

Repository files navigation

ARC-7 Logo

ARC-7

A Research-Backed, Multi-Agent Architectural Review Panel

Stop relying on a single LLM to validate your system architecture.


ARC-7 is a tool-agnostic, multi-agent system that convenes a panel of 7 highly specialized AI personas to conduct rigorous architectural reviews. By leveraging cognitive diversity, structured adversarial debate, and ensemble learning, ARC-7 produces enterprise-grade system validation that far exceeds the capabilities of any single foundation model.

🧠 The Science: Why ARC-7?

Recent advancements in AI research demonstrate that single-shot prompts to monolithic models suffer from "groupthink," sycophancy, and high hallucination rates when tackling complex, high-stakes system design.

ARC-7 solves this by implementing proven cognitive frameworks from AI research:

  • Drastic Reduction in Hallucinations: By utilizing an adversarial review dynamic, claims made by one agent are inherently fact-checked by the others. The Context Master orchestrator synthesizes and verifies all findings, dropping ungrounded hallucination rates to near-zero.
  • Ensemble Learning & Cognitive Diversity: ARC-7 doesn't just use 7 prompts—it is designed to run on a Mixture of Models (e.g., Claude 3.5 Sonnet, GPT-4o, Gemini 1.5 Pro). Because different foundation models possess different training distributions and latent biases, combining them creates a cognitive mesh that catches edge cases a single model would blindly miss.
  • Mitigating Sycophancy (The "Yes-Man" Problem): Standard LLMs are trained to agree with the user. ARC-7 explicitly injects a Naysayer (designed to find failure modes) and a Security Sentinel (with absolute veto power over insecure designs) to ensure brutal, mathematically honest feedback.
  • Structured Conflict Resolution: When models disagree on a fundamental architectural choice, ARC-7 initiates a blind voting protocol. If the panel remains split, it formally recommends a prototype spike, preventing premature optimization.

🎭 The Panel Members

Every persona is engineered with a hyper-specific cognitive profile and is mapped to the foundation model best suited for that type of reasoning.

Role Focus & Responsibility Recommended Model Matrix
👑 Context Master Orchestration, synthesis, conflict resolution, deduping gemini-3.1-pro (Massive Context)
🏛️ The Architect System design, API contracts, domain boundaries claude-3.5-sonnet (Deep Reasoning)
🛡️ Security Sentinel OWASP A01-A10, STRIDE, strict Enterprise Sec veto gpt-4o / o1-preview (Adversarial)
📈 Product Visionary ROI, user metrics, MVP scope creep prevention gpt-4o (Business Logic)
🎨 Creative Strategist UX innovation, pattern breaking, system simplification gpt-4o (Divergent Thinking)
The Optimizer Performance limits, cost control, parallelization gpt-4o (Algorithmic)
🛑 The Naysayer Reality checking, edge cases, finding hidden risks claude-3.5-sonnet (Skeptical Logic)

🚀 Capabilities & Review Modes

ARC-7 operates autonomously across three distinct execution modes depending on where you are in the SDLC:

  1. Conversation Review (/ARC-7)
    • Use Case: Early-stage brainstorming.
    • Action: Extracts proposals, decisions, and constraints from your current chat history and runs them through the panel to catch flaws before a single line of code or documentation is written.
  2. Document Review (/ARC-7 <document>)
    • Use Case: RFCs, Tech Specs, and System Design Docs.
    • Action: Reads a specific file or inline proposal and produces a formal, multi-perspective audit report.
  3. Codebase Review (/ARC-7 <github-url>)
    • Use Case: Legacy modernization, PR audits, or structural tech-debt analysis.
    • Action: Clones an entire repository to memory, maps the directory structure and tech stack, and evaluates systemic patterns, architectural drift, and component boundaries across the whole codebase.

🛠️ Tool-Agnostic Architecture

ARC-7 is strictly designed to be tool and provider agnostic. It will run on opencode, Aider, custom MCP servers, LangChain, or any framework that supports markdown-based system prompts and skills.

Directory Structure

  • skills/ARC-7/ — The master orchestration logic, rules of engagement, and prompt injection sequence.
  • agents/ARC-7/ — The individual persona definitions (system prompts, constraints, output schemas).
  • commands/ — UI slash-command definitions.

Provider-Agnostic Model Routing

To prevent API crashes across different execution environments, models are declared abstractly as recommended_model in the YAML frontmatter of each agent file.

To map these recommended models to your specific AI provider (OpenAI, Anthropic, GitHub Copilot, OpenRouter), see the included model-mappings.json and copy the relevant configuration into your tool's global config file (e.g., ~/.config/opencode/opencode.json).


🏁 Installation & Quick Start

ARC-7 includes cross-platform installation scripts that automatically route the core files into your specific agent tool's hidden configuration directories.

1. Run the installer:

# Mac/Linux
./install.sh --target .opencode 

# Windows (PowerShell)
.\install.ps1 -Target .agents 

(By default, this performs a safe file copy. If you are developing ARC-7 locally, you can pass --mode symlink to create live symlinks instead).

2. Configure your models (Fixes ProviderModelNotFoundError): Run the configuration script to automatically read your provider from model-mappings.json and insert the correct mapping block directly into your tool's global config file (e.g., ~/.config/opencode/opencode.json). It will only insert mappings that do not already exist. If you do not run this step, the Task tool will fail with a ProviderModelNotFoundError because the orchestrator will not know which model or provider to use when spawning the 7 sub-agents.

# Mac/Linux
./configure_models.sh --provider github_copilot
# Other options: openai_direct, openai_anthropic_direct, anthropic_direct, google_ai_studio, azure_openai, ollama, groq, openrouter, aws_bedrock, opencode_zen, factory_ai, together_ai

# Windows (PowerShell)
.\configure_models.ps1 -Provider github_copilot

3. Invoke the panel:

/ARC-7                # Audit the current conversation
/ARC-7 docs/rfc.md    # Audit a specific architectural document
/ARC-7 https://github.com/org/repo/tree/main   # Audit a full codebase
/ARC-7 help           # View full usage instructions

Don't build in an echo chamber. Let ARC-7 tear your architecture apart before production does.


📚 Academic Grounding & References

The architectural principles behind ARC-7's multi-agent workflow are grounded in recent machine learning and cognitive science research. By implementing these peer-reviewed methodologies, ARC-7 fundamentally outperforms single-prompt/single-model inference:

  1. Multi-Agent Debate & Hallucination Reduction
    • Improving Factuality and Reasoning in Language Models through Multiagent Debate (Du et al., 2023) - arXiv:2305.14325. Demonstrates how multiple agents debating and reviewing each other's responses significantly mitigates hallucination and improves reasoning accuracy over single-agent systems.
  2. Mitigating Sycophancy (The "Yes-Man" Problem)
    • Towards Understanding Sycophancy in Language Models (Sharma et al., 2023) - arXiv:2310.13601. Details the pervasive issue of LLMs tailoring their responses to agree with the user's unstated beliefs. ARC-7 actively combats this via hardcoded adversarial constraints (The Naysayer, Security Sentinel).
  3. Cognitive Diversity & Ensemble Learning in LLMs
    • More Agents Is All You Need (Li et al., 2024) - arXiv:2402.05120. Provides empirical evidence that scaling the number of diverse, specialized agents through ensemble frameworks mathematically increases the probability of discovering edge cases and arriving at optimal solutions.
  4. Role-Playing & Persona Alignment
    • CAMEL: Communicative Agents for "Mind" Exploration of Large Language Model Society (Li et al., 2023) - arXiv:2303.17760. Highlights the efficacy of strict persona-based prompting and cross-agent communication to solve complex, multi-step engineering tasks.

About

Agentic Architecture Review Panel skill

Resources

License

Stars

Watchers

Forks

Packages

 
 
 

Contributors