MS-Agent Documentation

Code Genesis is a production-ready, open-source multi-agent framework designed for end-to-end project generation. It utilizes a topology-aware code synthesis approach, meaning it uses dependency-driven scheduling to prevent hallucinated references and enable parallel generation. The framework follows a holistic engineering pipeline that includes requirement analysis, architecture design, implementation, and deployment.

Key technical features include:

• Topology-Aware Code Synthesis: Uses dependency-driven scheduling to ensure correct file ordering.
• LSP-Integrated Validation: Incorporates real-time Language Server Protocol (LSP) feedback during the generation process to ensure code correctness.
• Multi-Agent Workflow: Employs specialized agents for different stages of the software development lifecycle (SDLC).

Code Genesis offers two distinct pipeline modes depending on your needs:

1. Standard Workflow (Production-Grade): A 7-stage process (User Story → Architect → File Design → File Order → Install → Coding → Refine) designed for complex, high-quality applications with explicit architectural planning and dependency-driven generation.
2. Simple Workflow (Rapid Prototyping): A 4-stage condensed process (Orchestrator → Install → Coding → Refine) optimized for speed and lightweight projects or quick iterations.

The FinResearch workflow is orchestrated via a Directed Acyclic Graph (DAG) of five specialized agents. Key configuration and logic files include:

• workflow.yaml: The main entry point defining the execution process.
• agent.yaml files: Individual configurations for the Orchestrator, Searcher, Collector, Analyst, and Aggregator agents.
• conf.yaml: Configuration for search engines (e.g., Exa, SerpAPI) used by the Searcher agent.
• callbacks/: Modules for handling agent outputs (e.g., orchestrator_callback.py saves plans, aggregator_callback.py saves final reports).
• tools/: Contains build_jupyter_image.sh for Docker sandbox setup and principle_skill.py for analytical frameworks.
• searcher.py: Triggers ms-agent/projects/deep_research for sentiment searches.
• aggregator.py: Merges sentiment and quantitative analysis results.

The standard workflow follows a 7-stage pipeline defined in workflow.yaml:

1. User Story Agent: Parses requirements into structured stories.
2. Architect Agent: Selects tech stack and defines architecture.
3. File Design Agent: Generates physical file structure.
4. File Order Agent: Constructs dependency DAG and topological sort.
5. Install Agent: Bootstraps environment and resolves dependencies.
6. Coding Agent: Synthesizes code with LSP validation.
7. Refine Agent: Performs runtime validation, bug fixing, and automated deployment.

The MEMORY.md file serves as the long-term memory storage for the nanobot agent. It is designed to persist important information across different sessions. The file is structured into four main sections:

• User Information: Stores important facts about the user.
• Preferences: Stores user preferences learned by the agent over time.
• Project Context: Stores information regarding ongoing projects.
• Important Notes: Stores miscellaneous things the agent should remember.

Note: This file is automatically updated by nanobot when the agent identifies information that should be remembered.

MS-Agent offers several core capabilities:

• Multi-Agent for general purpose: Chat with agents featuring tool-calling capabilities based on MCP.
• Deep Research: Enables advanced autonomous exploration and complex task execution.
• Code Generation: Supports code generation tasks with artifacts.
• Short Video Generation: Supports video generation of approximately 5 minutes.
• Agent Skills: A knowledge-driven skill system providing reusable procedural knowledge via standard tool integration. It features multi-source loading, progressive disclosure, and runtime self-evolution.
• WebUI: A modern web interface for agent interaction using real-time WebSocket communication.
• Lightweight and Extensible: Designed for easy customization and extension.

The Simple Workflow is a condensed pipeline for rapid prototyping:

• Orchestrator Agent: A single reasoning step that handles requirement analysis, architecture design, and file planning.
• Install Agent: Handles dependency resolution and environment setup.
• Coding Agent: Performs direct code generation with integrated file ordering.
• Refine Agent: Handles final validation and deployment.

The LLMAgent class is the core component in MS-Agent responsible for conversational capabilities and tool invocation. It manages the execution loop, including LLM calls, tool usage, message management, and RAG.

Execution Loop Logic:

1. Registers configured callbacks.
2. Initializes LLM, tools, message management, and RAG.
3. Reads message history cache.
4. Prepares messages (RAG queries, plans, etc.).
5. Enters loop: attempts to compress messages, prepares tool list, calls LLM, and executes tools based on results.

Loop Termination Conditions:

• The model provides a reply without any tool calls.
• The number of rounds reaches max_chat_round specified in the configuration.

MS-Agent uses a three-level progressive disclosure mechanism to manage context window costs and detail levels:

1. L1 (Automatic): The system prompt includes a lightweight index (Name + one-line description) for all enabled skills (~30 tokens per skill).
2. L2 (On Demand): The model calls the skill_view(skill_id) tool to load the full SKILL.md body.
3. L3 (On Demand): The model uses skill_view with a file_path to access referenced scripts, templates, or documentation files.