Agentic Workflow: A Multi-Agent Framework for Claude Code Enabling Hierarchical Review and Skill Evolution

Agentic Workflow is a multi-agent framework specifically designed for Claude Code, aiming to solve the quality and trust issues in AI programming. Its core features include S/M/L tiered acceptance strategies, adversarial review, cross-model second opinions, skill evolution mechanisms, and explicitly positions humans as the final arbiters. This project is maintained by AgentShekel, with source code hosted on GitHub: https://github.com/AgentShekel/agentic-workflow, and was released on May 29, 2026. The framework establishes layered quality safeguards through multi-agent collaboration, enhancing output reliability while maintaining AI programming efficiency.

AI coding assistants are rapidly improving their capabilities, capable of completing tasks such as natural language-to-code conversion, code refactoring, multi-step development tasks, testing and fixing. However, erroneous code may lead to production failures, security vulnerabilities, or data loss. Traditional human-in-the-loop solutions are safe but inefficient. Agentic Workflow explores a third path: establishing layered quality safeguards through multi-agent collaboration, where only high-risk changes require human intervention, balancing efficiency and security.

The framework simulates the division of labor in a software development team, including the following roles:

1. Executor: Assumed by Claude Code, responsible for understanding requirements, analyzing context, formulating plans, and implementing changes.
2. Separation of Acceptor and Optimizer: The acceptor conducts initial output review (tiered strategy), while the optimizer fixes issues targetedly, reducing self-cognition bias.
3. Adversarial Reviewer: Proactively looks for code issues (security vulnerabilities, logical errors, etc.) in an isolated environment, simulating security audits.
4. Cross-Model Second Opinion: Obtains independent evaluations from other model families (e.g., GPT series) via Codex MCP to reduce blind spots of a single model.

Tiered Acceptance Strategy (S/M/L Tiering)

• S Tier: Minor changes (single-line modifications, document updates) can be automatically accepted after passing automated tests.
• M Tier: Medium changes (function refactoring, module adjustments) require acceptor review and may trigger optimization cycles.
• L Tier: Major changes (architecture adjustments, dependency upgrades) require adversarial review + cross-model validation + human arbitration.

Skill Evolution Mechanism

Drawing on the SkillOpt concept: Extract reusable patterns from successful tasks → maintain a structured skill library → dynamically apply them when executing new tasks → continuously optimize based on feedback to achieve framework capability iteration.

Event Ledger and Observability

Records complete task trajectories, decision reasons, review disagreements, and other logs, supporting auditing, debugging, and real-time monitoring. Notifies administrators when anomalies occur.

Use Cases

• Enterprise codebase maintenance: Secure automated refactoring and updates.
• Open source project contributions: Automatically handle issues, generate PRs, and ensure quality.
• Security-sensitive development: Multi-layered security guarantees for sensitive modules (authentication, payment).
• Team knowledge precipitation: Encode best practices into agent skills.

Human Roles

The framework retains human final authority: L-tier changes require manual approval; agents can request human intervention; humans can override automatic decisions; key configurations need manual confirmation.

Agentic Workflow represents an important direction for AI programming assistants to evolve into production-grade tools: shifting from single-agent capability demonstration to a reliable system of multi-agent collaboration. Through hierarchical review, adversarial evaluation, cross-model validation, and skill evolution, it balances efficiency and quality. For teams adopting AI programming, this framework provides a reference architecture, emphasizing that the stronger the AI capability, the more critical the governance and review mechanisms become.