Claude Agent Dispatch: A Tag-Driven Autonomous Workflow System for Claude Code

Claude Agent Dispatch is a reusable agent dispatch system that enables a fully automated workflow from issue triage to code implementation and PR creation via GitHub Actions and a tag-driven state machine. It addresses the engineering challenges of scaling Claude Code capabilities from individual development to team collaboration, and converting ad-hoc sessions into automated processes. Key features include a tag-driven state machine, two-stage approval, worktree isolation, multi-layered security mechanisms, etc., representing the evolution of AI-assisted development from tools to workflows.

AI coding assistants have evolved from simple code completion tools to agents capable of executing complex tasks. Claude Code, as a command-line AI assistant launched by Anthropic, can collaborate with developers in the terminal to understand codebases, execute commands, and modify files. However, scaling its capabilities to team collaboration and automated workflows still requires significant engineering effort, which is the problem the Claude Agent Dispatch project aims to solve.

Claude Agent Dispatch is a reusable agent dispatch system that encapsulates Claude Code capabilities into tag-driven GitHub Actions workflows. Key features include:

• Tag-driven state machine: Control workflows via GitHub tags
• Two-stage approval: Manual review during the planning phase to balance security and efficiency
• Worktree isolation: Support concurrent processing of multiple issues
• Built-in security mechanisms: Circuit breakers, tool restrictions, concurrency control, etc.
• 52 BATS tests: Ensure system reliability

The system workflow is triggered by GitHub Issue tags, with a complete lifecycle: agent:triage → agent:plan-review → agent:plan-approved → agent:in-progress → agent:pr-open.

• Triage Phase: The agent reads the CLAUDE.md file, explores the codebase, and adds the agent:needs-info tag to ask for more information if required.
• Plan Review Phase: Formulate a detailed implementation plan, post a comment, and mark it as agent:plan-review to wait for manual review.
• In Progress Phase: After the plan is approved, follow TDD principles—commit code, run tests in each cycle, then push the branch and create a PR upon completion.
• PR Revision Phase: Read feedback from PR comments, automatically push fixes, and repeat until the PR is accepted.

To ensure the safety of autonomous agent operations, the system is designed with multi-layered protection:

• Circuit breaker mechanism: Maximum 8 bot comments per Issue per hour (configurable) to prevent infinite loops.
• Tool restrictions: Only explicitly allowed tools are permitted; dangerous operations like sudo and network access are disabled, with fine-grained control via AGENT_ALLOWED_TOOLS.
• Actor filtering: Bot's own operations do not trigger workflows to avoid loops.
• Concurrency control: Only one task runs per Issue/PR at a time.
• Timeout protection: Configurable task timeout (default 3600 seconds) to prevent resource hogging.

The system offers two deployment modes:

• Reference Mode: Workflow files call upstream reusable workflows to get automatic updates, suitable for out-of-the-box use.
• Standalone Mode: All scripts and prompts are copied to the project's .agent-dispatch/ directory for full control without dependencies, ideal for deep customization. Adaptation mechanisms include:
• CLAUDE.md: A project convention file read by the agent before handling Issues to ensure compliance with specifications.
• Custom prompts: Override default prompts via AGENT_PROMPT to add project-specific instructions.
• Test gate: Set AGENT_TEST_COMMAND to require the agent to pass tests before creating a PR.

Anthropic's official claude-code-action is a lightweight GitHub Action that responds to @claude mentions or simple PR interactions; Claude Agent Dispatch, on the other hand, is a complete autonomous agent lifecycle system with features like a tag-driven state machine, two-stage process, concurrent processing, and multi-layered security mechanisms. The two complement each other: the former is suitable for quick interactions, while the latter is ideal for full automated workflows.

Practical Recommendations:

1. Start with CLAUDE.md to organize project conventions;
2. Conduct small-scale pilots and select appropriate Issue types;
3. Emphasize plan reviews to balance security and efficiency;
4. Set up test gates to prevent code regression;
5. Continuously optimize prompts to adjust agent behavior.

Project Significance: Represents an important direction of AI-assisted development from tools to workflows, demonstrating a new mode of human-AI collaboration—AI handles tedious implementation tasks, while humans focus on requirement clarification, plan review, and quality control, helping teams improve efficiency and reduce repetitive work.