Agent Concepts
At its core, GitAgent Workbench manages an AgentWorkspace — a structured configuration that defines everything an AI agent needs to operate:
-
Manifest (
agent.yaml) — Metadata: name, version, description, dependencies, compliance settings -
Soul (
SOUL.md) — The agent’s core identity, personality, and mission -
Rules (
RULES.md) — Behavioral boundaries and operational constraints -
Prompt (
PROMPT.md) — System prompt for non-minimal structures - Skills — Reusable capability modules with their own instructions, references, and templates
- Tools — MCP-compatible function definitions the agent can call
- Workflows — Multi-step orchestration patterns for complex tasks
- Knowledge — Documents the agent can reference at runtime
The application supports 9 structure types that determine which components are generated: minimal, standard, full, data-analyst, web-scraper, researcher, inheritance, multi-repo, and monorepo.
Prerequisites
- Node.js 20+ and npm
- An API key for at least one supported AI provider:
- Anthropic (
ANTHROPIC_API_KEY) - OpenAI (
OPENAI_API_KEY) - Google Gemini (
GOOGLE_API_KEYorGEMINI_API_KEY) - Mistral (
MISTRAL_API_KEY) - Groq (
GROQ_API_KEY) - Ollama (local, set
OLLAMA_BASE_URL) - OpenRouter (
OPENROUTER_API_KEY)
- Anthropic (
Quick Start
# Clone the repository
git clone <repo-url>
cd gitagent-workbench
# Install dependencies
npm install
# Start the development server
npm run dev
Open http://localhost:3000 in your browser.
Creating Your First Agent
- Enter the Wizard — Click “Create New Agent” to start the configuration flow
- Identity — Name your agent, set its version and description
- Capabilities — Define skills, tools, and workflows
- Model — Select an AI provider and model for generation
- Compliance — Set risk tier, supervision mode, and review cadence
-
Structure — Choose the complexity level (start with
standard) - Review — Confirm your configuration
- Generate — The AI pipeline creates all agent files
- Export — Download the complete agent package as a ZIP
Key Concepts
Skills vs Tools
Skills are high-level capabilities with instructions, references, and templates. They tell the agent how to do something. Example: “Code Review” skill with guidelines for reviewing pull requests.
Tools are low-level function definitions (MCP-compatible) that the agent can call. They tell the agent what it can execute. Example: “read_file” tool that reads a file from disk.
The Generation Pipeline
When you click “Generate”, the orchestrator runs a sequential pipeline of up to 12 steps, conditionally executed based on your structure type:
- Sanitize inputs (truncate oversized context)
- Generate YAML manifest
- Define agent soul
- Generate instructions (rules, prompt, duties)
- Configure runtime
- Generate skills
- Draft knowledge docs
- Define tool schemas
- Configure sub-agents (if any)
- Plan workflows (full structure only)
- Generate examples (full structure only)
- Final validation
Export Format
The exported ZIP contains a complete, valid agent package:
agent.yaml # Manifest with dependencies and compliance
SOUL.md # Agent identity
RULES.md # Behavioral rules
PROMPT.md # System prompt (non-minimal)
skills/ # Skill directories with SKILL.md files
tools/ # Tool YAML definitions (MCP-compatible input_schema)
workflows/ # Workflow YAML files
knowledge/ # Knowledge documents with index
memory/ # Memory configuration
config/ # Runtime configuration
Next Steps
- Custom Tools — Learn how to define and use tools in your agents
- Agent Patterns — Best practices for agent design
- API Reference — Server API endpoint documentation
- Architecture — System overview and data flow
- Troubleshooting — Common issues and solutions