An agentic coding framework for Claude Code, GitHub Copilot CLI, and Microsoft Amplifier, featuring a 23-step structured workflow, 37 specialized agents, persistent knowledge graph memory, and L1–L12 progressive evaluation.
amplihack is an agentic coding framework designed for Claude Code, GitHub Copilot CLI, and Microsoft Amplifier. Its core philosophy elevates LLM coding capabilities from single-turn conversations to orchestrated, measurable, and self-improving systematic engineering workflows.
The framework includes four built-in workflows: DEFAULT_WORKFLOW (23-step systematic development process), INVESTIGATION_WORKFLOW (6-phase knowledge mining), Q&A_WORKFLOW (3-step lightweight Q&A), and OPS_WORKFLOW (1-step operations management). All workflows are enforced via the YAML Recipe Runner in a code-mandated manner—models cannot skip or bypass steps. Workflows are customizable through configuration file editing.
The agent system comprises 37 specialized AI agents (architect, builder, reviewer, tester, security, etc.), supporting Goal-Seeking Agents (with memory, evaluation, and self-improvement capabilities) and Session-to-Agent (converting interactive sessions into reusable agents). The skill system provides 85+ skills with context-based automatic activation.
The memory and knowledge system uses the Kuzu embedded graph database for cross-session persistent memory, paired with the independent amplihack-memory-lib providing a CognitiveMemory 6-type system. For code intelligence, vendored Blarify is integrated with Tree-sitter parsing support for 8 languages: Python, JavaScript, TypeScript, C#, Go, Java, PHP, and Ruby.
The quality assurance system includes quality gates (philosophical compliance, test coverage, code standards), an L1–L12 progressive evaluation system (with long-term memory testing and self-improvement), and a Gherkin BDD expert skill (26% accuracy improvement over natural language descriptions).
The orchestration layer uses /dev as the primary entry command, automatically classifying tasks, detecting parallel workflows, and orchestrating execution. It supports multiple independent features developed in parallel within /tmp clones. Auto Mode enables autonomous agentic loops, and Fleet Management supports multi-VM orchestration. The quick fix command /fix <pattern> rapidly resolves common errors in imports, CI, tests, and config.
Technically, Python serves as the core body handling orchestration logic, workflow engine, and agent definitions, while a Rust runtime (gradually migrating) handles the recipe runner and hook engine for performance. The build system is based on setuptools with custom build_hooks.py, using uv as the package manager. Platform support includes full macOS/Linux/WSL and partial Windows native. Prerequisites are Python 3.11+, Node.js 18+, git, and uv.
Unconfirmed information: Open-source license is not yet specified (no LICENSE file in repository); the official documentation site URL is not displayed in the README.