Agent Foundry: A Multi-Agent Development Workflow Framework for Claude Code CLI

Agent Foundry is an open-source multi-agent development workflow framework built for Claude Code CLI. It includes over 150 skills and 5 professional agents, enabling a complete AI-assisted development process from design and implementation to review through multi-model orchestration. Its goal is to address the limitations of single-model interactions and improve development quality and efficiency. The project is maintained by community developers, adopts clear role division and collaboration processes, and supports features like persistent task management and continuous review.

With the in-depth application of large language models in software development, the limitations of single-model interactions have gradually become apparent. Claude Code CLI, as Anthropic's command-line programming assistant, has strong code understanding and generation capabilities, but building more complex and systematic AI-assisted workflows has become a direction of community exploration. Agent Foundry emerged as a result—it is not just a collection of skills but a complete multi-agent orchestration framework maintained by community developer joogy06. Through clear role division and collaboration processes, it allows multi-agents to leverage their strengths in each stage of development to achieve higher-quality outputs.

Agent Foundry follows the 'routing-design-implementation-review' pipeline model, with five agents each taking their own responsibilities:

• pa (Persistent Agent)：Task routing and state management. As the system's entry coordination center, it understands user intent, routes tasks, maintains cross-session states, and solves the problem of context loss.
• forge: Design exploration and multi-model challenge. It uses three models (Claude, Codex, Gemini) to question and complement each other, producing well-argued design documents and reducing single-model bias.
• bob: Autonomous implementation executor. It decomposes work packages according to approved designs, plans execution sequences, follows the 'design-driven development' concept, and avoids arbitrary execution.
• alf: Evolution and improvement review. It regularly reviews code and skills, identifies issues, and emphasizes continuous evolution to ensure the long-term health of the project.
• wiki: Knowledge base builder. It organizes structured knowledge base pages, and the 'compile once, read many times' approach reduces LLM processing costs.

Agent Foundry includes over 150 domain skills covering software engineering, DevOps, and other fields. It uses front-end metadata description, allowing Claude Code to load them automatically as needed. Featured skills include:

• publish-to-github: A secure release process that performs security checks before pushing;
• visual-companion: Browser-based collaborative review of prototype diagrams to compensate for the limitations of plain text. Quick start: Run python3 bootstrap-environment.py to set up the environment with one click, including 10 idempotent steps (such as installing skills to ~/.claude/, configuring services, etc.). Pre-commit hooks scan for sensitive information before git push—high-severity issues block the push, while medium and low-severity issues issue warnings.

The core innovation is multi-model orchestration: forge calls three models (Claude, Codex, Gemini) simultaneously during the design phase to simulate human design reviews and avoid single-model fixed thinking. It supports three usage levels:

• Minimal level: Only Claude Code CLI, with access to over 150 skills and the wiki agent;
• Standard level: Add Codex/Gemini CLI MCP to enable multi-model review by forge, bob, and alf;
• Full level: Add pa-server and claude-in-chrome MCP to gain persistent task tracking and browser review capabilities.

Application scenarios include:

• Complex feature development: Multi-agent collaboration to oversee each stage;
• Codebase maintenance: alf regularly scans for parts that need refactoring or updates;
• Knowledge precipitation: wiki organizes structured knowledge bases to reduce onboarding costs;
• Security compliance: Pre-commit hooks and policy restrictions to maintain the security bottom line. Community ecosystem: Modular design—skills and agents follow interface specifications, allowing developers to create and share their own skill packages/agents. The documentation includes dependency instructions and troubleshooting guides, supporting Windows (PowerShell) and POSIX (Bash) systems.

Agent Foundry represents the evolutionary direction of AI-assisted development from single-model to multi-agent collaboration. Through clear division of labor, standardized processes, and multi-model verification, it introduces engineering quality control while maintaining AI flexibility. For Claude Code CLI users, it provides a systematic workflow path; for observers, it is a practical case of multi-agent orchestration. As LLM capabilities improve and multi-model collaboration matures, such frameworks will become an important bridge connecting human developers and AI, playing a more significant role in the software development field.