BreadBoard - Agentic Coding System

A streamlined, modular implementation of an agentic coding system that can generate, modify, and debug code autonomously. Works with gVisor template containers that are autoscaled by Ray.

Quick Start

Basic Usage

# Run a task with a specific config
python main.py agent_configs/test_enhanced_agent_v2.yaml -t "Create a fibonacci calculator in Python"

# Start interactive session
python main.py agent_configs/test_enhanced_agent_v2.yaml -i

# Use custom workspace
python main.py agent_configs/test_enhanced_agent_v2.yaml -w my_project -t "Build a web scraper"

Configuration

Agent configurations are stored in the agent_configs/ directory. Key configs:

• test_enhanced_agent_v2.yaml - Enhanced agent with multiple tool calling modes
• test_opencode_gpt5_nano_c_fs.yaml - Specialized for C filesystem tasks
• test_simple_native.yaml - Simple native tool calling setup

Architecture

Core Components

• agentic_coder_prototype/ - Main agent implementation

  • agent.py - Simplified interface
  • agent_llm_openai.py - Core LLM agent logic
  • provider_routing.py - Multi-provider support
  • provider_adapters.py - Provider-specific adapters
  • tool_calling/ - Tool calling system

• agent_configs/ - Agent configuration files

• breadboard/ - Sandboxing + LSP utilities

  • breadboard/sandbox_v2.py - Secure execution environment
  • breadboard/lsp_manager_v2.py - Language server integration

• implementations/ - Tool definitions and system prompts

• tui_skeleton/ - Claude Code-style TUI + CLI (breadboard binary)

Key Features

• Multi-Provider Support: OpenAI, Anthropic, OpenRouter
• Secure Sandboxing: Isolated execution environments
• LSP Integration: Language server support for enhanced code intelligence
• Flexible Tool Calling: Multiple syntax formats and execution modes
• Modular Design: Clean separation of concerns

Development

Project Structure

├── main.py                     # CLI entry point
├── agentic_coder_prototype/    # Core agent system
├── agent_configs/              # Agent configurations
├── breadboard/                 # Sandbox + LSP utilities
├── implementations/            # Tools and prompts
├── tui_skeleton/               # TUI + CLI (Node/Ink)
├── tests/                     # Test suite
└── misc/                      # Non-essential files

Testing

# Run specific tests
python -m pytest tests/test_agent_session.py -v

# Run all tests
python -m pytest tests/ -v

Phase 9 Utilities

• Surface manifest smoke check:

  python scripts/phase9_surface_smoke_check.py --run-dir <logging_run_dir>

• TreeRunner scaffold config snippet:

  ctrees:
    runner:
      enabled: true
      branches: 2

Configuration

Agent behavior is controlled through YAML configuration files. Key parameters:

• model: LLM model to use (e.g., "openrouter/openai/gpt-4")
• max_iterations: Maximum agent loop iterations
• tool_prompt_mode: Tool calling syntax preference
• tools_file: Path to tool definitions

Examples

Creating a Simple Calculator

python main.py agent_configs/test_enhanced_agent_v2.yaml -t "Create a calculator.py that can add, subtract, multiply, and divide two numbers with error handling"

Building a Web API

python main.py agent_configs/test_enhanced_agent_v2.yaml -t "Create a FastAPI web service with endpoints for user management (create, read, update, delete users)"

Interactive Development

python main.py agent_configs/test_enhanced_agent_v2.yaml -i
> Create a data processing script
> Add unit tests for the script  
> Fix any bugs found in testing
> exit

License

This project is experimental research software. Use at your own risk.