Design Arena Agent Runner

A model-agnostic, framework-agnostic agent harness that enables autonomous AI agents across any LLM provider.

Beta Version - Currently in active development.

Features

• Turn Any Model Into an Agent: Configure any model as a coding agents
• Flexible Configuration: Set up agent behavior, tools, and constraints however you need
• CLI or Python API: Use as a command-line tool or integrate into your applications
• Powerful Tooling: File operations, code editing, search, bash execution
• Smart Context Management: Automatic context window handling and compaction
• Workspace Isolation: Secure file operations with command validation

Installation

Prerequisites

Python:

• Python 3.11 or higher

Install Agent Runner

Option 1: From PyPI (recommended)

pip install agent-runner

Option 2: From source

git clone https://github.com/Design-Arena/agent-runner.git
cd agent-runner
pip install -e .

With optional dependencies:

# Development tools (testing, linting)
pip install agent-runner[dev]

# Optional tools (screenshot, patch tools)
pip install agent-runner[tools]

# Additional providers (Mistral)
pip install agent-runner[providers]

# Everything
pip install agent-runner[all]

For screenshot tool (optional):

pip install agent-runner[tools]
playwright install chromium

For search functionality (recommended):

Install ripgrep for fast code search:

# macOS
brew install ripgrep

# Ubuntu/Debian
sudo apt-get install ripgrep

Configure API Keys

Set up your API keys as environment variables. You only need keys for the providers you'll use.

Create a .env file in your project directory:

OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...
GOOGLE_API_KEY=...
MISTRAL_API_KEY=...
XAI_API_KEY=...
KIMI_API_KEY=...
ZAI_API_KEY=...

Or set environment variables directly:

export OPENAI_API_KEY="sk-..."
export ANTHROPIC_API_KEY="sk-ant-..."
# etc.

Quick Start

Basic Usage

Run a single task:

agentrunner run "Create a Python calculator with add and multiply functions"
agentrunner run "Build a FastAPI todo app" --verbose  # See live events

Interactive chat:

agentrunner chat

Review code (read-only, no modifications):

agentrunner review .
agentrunner review src/main.py

Model management:

agentrunner models  # List all available models
agentrunner run "Build a REST API" --model claude-sonnet-4-5-20250929
agentrunner run "Analyze this code" --model gemini-2.5-pro

Session management:

agentrunner sessions list              # List all sessions
agentrunner sessions show <id>         # View session details
agentrunner sessions show <id> --events  # View with full event log
agentrunner sessions delete <id>       # Delete a session

Note: There are two types of sessions:

• sessions (plural) - CLI session history with full event traces
• session (singular) - Agent conversation sessions that can be resumed

Configuration:

agentrunner config list              # List profiles
agentrunner config show              # Show current profile
agentrunner config set-default --model claude-sonnet-4-5-20250929  # Set default model

Usage from Python

Basic example:

import asyncio
import os
from agentrunner.core.config import AgentConfig
from agentrunner.core.factory import create_agent
from agentrunner.providers.base import ProviderConfig

async def main():
    # Set your API key
    os.environ['OPENAI_API_KEY'] = "your-api-key"
    
    # Configure agent behavior (orchestration settings)
    agent_config = AgentConfig(
        max_rounds=50,           # Maximum number of agent turns
        tool_timeout_s=120       # Timeout for tool execution
    )
    
    # Configure the LLM provider (model settings)
    provider_config = ProviderConfig(
        model="gpt-5-codex",     # Model to use
        temperature=0.7,         # Sampling temperature
        max_tokens=4096          # Max tokens in response
    )
    
    # Create the agent
    agent = create_agent(
        workspace_path=".",              # Working directory
        provider_config=provider_config,
        agent_config=agent_config,
        profile="default"                # Optional profile name
    )
    
    # Run a task
    result = await agent.process_message(
        "Create a REST API for user management with FastAPI"
    )
    
    print(f"Result: {result.content}")

if __name__ == "__main__":
    asyncio.run(main())

Using different providers:

# Using Claude Sonnet 4.5
os.environ['ANTHROPIC_API_KEY'] = "your-api-key"
provider_config = ProviderConfig(
    model="claude-sonnet-4-5-20250929",
    temperature=0.7
)

Available Tools

Agent Runner provides a comprehensive set of tools that work consistently across all LLM providers:

File Operations

• read_file - Read file contents
• write_file - Write content to files
• create_file - Create new files
• delete_file - Delete files
• edit_file - Edit files with search/replace
• multi_edit - Multiple edits in one operation
• insert_lines - Insert lines at specific positions
• batch_create_files - Create multiple files at once

Search

• grep - Fast code search using ripgrep

Execution & Project Management

• bash - Execute bash commands
• scaffold_project - Generate project templates
• clean_workspace - Clean up workspace directory

Media & Generation

• take_screenshot - Capture screenshots (requires Playwright)
• fetch_image - Fetch images from URLs
• generate_image - Generate AI images
• fetch_video - Fetch videos from URLs
• generate_video - Generate AI videos

Deployment

• deploy_to_vercel - Deploy to Vercel

Note: All file operations are restricted to the workspace directory. Bash commands are validated against a whitelist and require user confirmation for dangerous operations, but exercise caution.

Supported Providers

Agent Runner provides a unified interface for multiple LLM providers:

OpenAI • Anthropic • Google • xAI • Mistral • Moonshot AI (Kimi) • Z.AI

See src/agentrunner/providers/registry.py for the complete list of available models.

Security Note

Agent Runner executes code and commands within your specified workspace directory. While file operations are restricted to the workspace, the agent can modify/delete files within it and execute bash commands. Always use a dedicated directory for agent work, exercise caution when pointing it at sensitive directories, and use version control to track changes.

Configuration

Environment Variables

Setting defaults via CLI:

# Set default model
agentrunner config set-default --model claude-sonnet-4-5-20250929

# Set default temperature
agentrunner config set-default --temperature 0.8

# Set multiple defaults at once
agentrunner config set-default --model gpt-5-codex --temperature 0.7

This creates/updates a .env file in your current directory.

Or set them manually in .env file:

Provider defaults (model, temperature, etc.):

# Default model (if not specified with --model)
export AGENTRUNNER_MODEL="claude-sonnet-4-5-20250929"

# Default temperature (if not specified with --temperature)
export AGENTRUNNER_TEMPERATURE="0.7"

# Max tokens for responses (optional)
export AGENTRUNNER_MAX_TOKENS="4096"

Agent behavior:

# Maximum agentic loop iterations
export AGENTRUNNER_MAX_ROUNDS=100

# Tool execution timeout in seconds
export AGENTRUNNER_TOOL_TIMEOUT=180

Profile Configuration

Profiles control agent orchestration behavior. Create JSON files in ~/.agentrunner/profiles/:

Example: ~/.agentrunner/profiles/default.json

{
  "max_rounds": 50,
  "tool_timeout_s": 120,
  "response_buffer_tokens": 1000,
  "allow_streaming": true
}

Example: ~/.agentrunner/profiles/thorough.json (for complex tasks)

{
  "max_rounds": 100,
  "tool_timeout_s": 300
}

Use with: agentrunner run "task" --profile thorough

Note: Model selection (--model), temperature (--temperature), and token limits (--max-tokens) are specified via CLI flags, not in profiles.

Project Configuration

Create .agentrunner/config.json in your project to override profile settings per-project:

{
  "max_rounds": 100,
  "tool_timeout_s": 180
}

Usage Examples

Create a New Project

mkdir ~/my-portfolio && cd ~/my-portfolio
agentrunner run "Create a cyberpunk-themed portfolio website with neon accents, animated background, project showcase grid, and contact form using Next.js, shadcn/ui, and Tailwind"

Code Review & Refactoring

cd ~/my-project
agentrunner run "Review src/api.py for security issues and refactor to use async/await"
agentrunner review src/api.py

Interactive Development

agentrunner chat
> Add error handling to all API endpoints
> Write unit tests for the new error handlers
> Run the tests with pytest
> quit

Debugging

agentrunner run "The tests in test_api.py are failing. Investigate and fix the bugs"

Generate Images & Media

agentrunner run "Generate a futuristic cityscape at sunset with neon lights and flying cars, then set it as the hero section background image on my landing page"

Architecture

agentrunner/
├── core/           # Agent logic, session management, config
├── providers/      # LLM provider integrations
├── tools/          # Tool implementations
├── security/       # Command validation, sandboxing
└── cli/            # Command-line interface

Key Components:

• Agent: Orchestrates LLM interactions and tool execution
• Providers: Unified interface for LLM APIs
• Tools: Extensible tool system
• Session: Conversation history persistence
• Context Strategy: Context window management

Troubleshooting

Command Not Found

If agentrunner command is not found:

# Make sure you're in the virtual environment
source .venv/bin/activate  # macOS/Linux
# or
.venv\Scripts\activate  # Windows

# Reinstall if needed
pip install -e .

API Key Issues

# Verify API key is set
echo $OPENAI_API_KEY

# Source .env file if using one
cd /path/to/agentrunner
set -a && source .env && set +a

Session Management

CLI Session History:

# List all CLI sessions (with event traces)
agentrunner sessions list

# View session details and messages
agentrunner sessions show <session-id>

# View with full event log (tool calls, file changes, etc.)
agentrunner sessions show <session-id> --events

# Delete a CLI session
agentrunner sessions delete <session-id>

Agent Conversation Sessions:

# List agent sessions (resumable conversations)
agentrunner session list

# Load and resume a conversation
agentrunner session load <session-id>

# Delete an agent session
agentrunner session delete <session-id>

Permission Errors

Agent Runner restricts file operations to the workspace directory for safety. Make sure you're running from the correct directory and have write permissions. Always use a dedicated workspace directory, not your home or system directories.

Run Tests

# All tests
pytest

# With coverage
pytest --cov=agentrunner --cov-report=html

# Specific test
pytest tests/unit/test_agent.py

Contributing

Contributions welcome! See CONTRIBUTING.md for guidelines.

License

This project is licensed under the MIT License - see the LICENSE file for details.