A structured, production-ready multi-agent framework for coordinating specialized AI modes using clear contracts, one-tool-per-message execution, and traceable task flows.
- Supports multi-mode agent teams (Orchestrator, Architect, Planner, Code, Debug, etc.)
- Enforces scoped edits, deterministic workflows, and boomerang-style task returns
- Works across any compatible AI runtime or platform (not tied to a single vendor)
- Ships with reusable templates for modes, instructions, and slash-commands
🔗 Quick Links: templates/ · meet-the-team/ · slash-commands/ · AGENTS.md (if present)
Professional AI Team Management - Deploy specialized AI agents with enterprise-grade coordination, advanced prompt engineering, and systematic workflow automation for superior development outcomes.
- Structured multi-agent coordination using clear Orchestrator / Worker / Reviewer roles
- Boomerang-style task returns for traceable, auditable workflows
- Token-aware, minimal-diff editing patterns for safe automated changes
- Production-oriented architecture with explicit scopes, contracts, and documentation
- Extensible templates for modes, instructions, and slash-commands without locking into specific tools
Use with any agentic runtime that supports:
- Multiple modes / roles (e.g., Orchestrator, Planner, Code, Debug)
- One-tool-per-message execution
- Custom/system instructions per mode
- File-scoped, deterministic edits
Examples (non-exclusive):
- Roo / Responses-style runtimes
- Kilo Code
- Other IDE or API-based agents with similar capabilities
git clone https://github.com/Mnehmos/Advanced-Multi-Agent-AI-Framework.git
cd Advanced-Multi-Agent-AI-Framework- Global instructions and mode contracts (see
AGENTS.mdif present) - Mode templates in
templates/custom_modes.yaml - Shared instructions in
templates/custom-instructions-for-all-modes.md - Slash command designs in
slash-commands/
Use the sections below ("How to use this as a GitHub Template" and "Compatibility / Requirements") to wire these files into your environment and start orchestrating work via the Orchestrator/Worker/Reviewer pattern.
Use this repository as a starting point for your own multi-agent setup:
- In GitHub, click "Use this template" on the repository page.
- Create your new repository from this template.
- In your new repo, keep the structure of:
templates/custom_modes.yamltemplates/custom-instructions-for-all-modes.mdtemplates/enhance-prompt-template.mdslash-commands/meet-the-team/AGENTS.md(if present) to house global contracts.
- In your chosen AI platform or runtime:
- Load the "custom instructions for all modes" into the global/system instructions.
- Load or adapt
custom_modes.yamlinto the platform's mode/multi-agent configuration. - Optionally register slash-commands based on
slash-commands/to standardize workflows.
- Start runs with an Orchestrator-style mode that:
- Decomposes work into subtasks,
- Assigns Workers with scoped file patterns,
- Routes results through a Reviewer when needed.
This keeps the framework portable while preserving the core coordination patterns.
This framework is designed to be environment-agnostic. Any runtime is compatible if it supports:
- One-tool-per-message or equivalent atomic tool execution
- Multiple modes / roles with distinct instructions
- Ability to enforce:
- Workspace paths and file pattern scopes
- Deterministic, inspectable steps
- Support for structured, JSON-like "boomerang" task return payloads
Examples:
- Roo-style / Responses-style orchestrated environments
- Kilo Code (using custom modes and project/global instructions)
- Custom in-house agentic runtimes wired to follow the same contracts
| Mode | Specialization | Advanced Techniques |
|---|---|---|
| 🔄 Orchestrator | Project Management & Task Delegation | workflow-template-prompting, boomerang-task-delegation |
| 🏗️ Architect | System Design & Architecture | visual-documentation-generation, tree-of-thoughts |
| 📅 Planner | Product Planning & Requirements | user-story-prompting, stakeholder-perspective-analysis |
| Mode | Specialization | Advanced Techniques |
|---|---|---|
| ⚒️ Builder | Software Development & Testing | code-generation-agents, test-based-iterative-flow |
| 💻 Code | Advanced Coding & Optimization | 'modular-code-generation, (https://github.com/chonghin33/lcm-1.13-whitepaper)' 'language-construct-modeling` |
| 🔒 Guardian | Infrastructure & CI/CD | automated-development-workflows, semantic-guardrails |
| Mode | Specialization | Advanced Techniques |
|---|---|---|
| ❓ Ask | Information Discovery | rag, iterative-retrieval-augmentation |
| 🔎 Deep Research | Comprehensive Analysis | multi-perspective-analysis, systematic-literature-review |
| 🔬 Deep Scope | Issue Analysis & Scoping | codebase-impact-mapping, hypothetical-scenario-modeling |
| Mode | Specialization | Advanced Techniques |
|---|---|---|
| 🐛 Debug | Technical Diagnostics | five-whys-prompting, chain-of-verification |
| 📁 Memory | Knowledge Management | semantic-clustering, knowledge-graph-construction |
- Structured multi-agent assistance for large codebases
- Automated, scope-safe refactors and reviews
- Task-mapped execution with clear success criteria
- Multi-mode assistants (Ask, Plan, Code, Debug) behind a single interface
- Repeatable workflows via shared slash-commands
- Auditable changes for safety and compliance
- CI/CD-aligned agent workflows
- Documentation and knowledge base automation
- Safe infrastructure and config updates via scoped Workers
Specification → Pseudocode → Architecture → Refinement → Completion
- Task Creation - Orchestrator generates structured tasks from project requirements
- Specialist Assignment - Tasks delegated to the most appropriate mode/agent
- Scoped Execution - Workers operate only within assigned workspace paths and file patterns
- Quality Integration - Results are validated (optionally via Reviewer) and merged back
- Iterative Improvement - Feedback loops refine instructions, scopes, and contracts
- Context window utilization kept below 40%
- Cognitive primitive optimization (start small, scale up)
- Specialized mode selection for minimal resource usage
- "Scalpel, not Hammer" resource management philosophy
- Structured task validation and success criteria
- Cross-mode verification and error checking
- Comprehensive documentation and traceability
- Automated workflow optimization
- Modular architecture supporting team expansion
- Customizable prompt engineering technique integration
- Enterprise workflow pattern implementation
- Professional project management capabilities
Detailed documentation for each AI specialist:
# Project: Advanced AI System Development
## Phase 1: Architecture Planning
- [ ] **Task 1.1**: System design and architecture planning
- **Agent**: Architect
- **Dependencies**: None
- **Outputs**: [architecture_diagram.md, technical_specifications.md]
- **Validation**: Architecture review completed with stakeholder approval
- **Human Checkpoint**: YES
- **Scope**: Complete system architecture design using visual-documentation-generation and tree-of-thoughts techniques- Structured documentation for audit trails
- Role-based task assignment and validation
- Quality gates and automated verification
- Professional workflow management
- GitHub integration for issue and PR management
- CI/CD pipeline automation
- Knowledge management system integration
- Custom prompt engineering technique deployment
- Comprehensive error handling and debugging
- Performance monitoring and optimization
- Documentation generation and maintenance
- Continuous improvement through feedback integration
- ⭐ Star this repository if you find the framework useful.
- 🛠️ Open issues or pull requests to refine contracts, docs, or examples.
- 🔗 Downstream users: keep this README neutral and add platform-specific details in your own overlays.
- Additional mode templates and role contracts
- More end-to-end examples for different runtimes
- Extended slash-command libraries for common workflows
MIT License - Open source framework for professional and commercial use.
- SPARC Framework development community
- Multi-agent AI research contributors
- Kilo Code platform development team
- Advanced prompt engineering research community
- Framework users providing feedback and improvements
- Vincent Shing Hin Chong for their work into Language Construct Modeling | https://osf.io/q6cyp/
- 20+ research papers sources listed here: https://mnehmos.github.io/Prompt-Engineering/sources.html
This repository provides a template-friendly, platform-agnostic entrypoint for:
- Defining clear multi-agent roles and contracts
- Running deterministic, auditable, one-tool-per-message workflows
- Using reusable templates for modes, shared instructions, and slash-commands
Start from this template, connect it to your preferred agent runtime, and coordinate complex work through the Orchestrator/Worker/Reviewer pattern with scoped, reviewable changes.