Self-Correcting AI Agent

A secure runtime for self-correcting AI agents with Docker sandboxing.

v1.0.0 Released 92% Success Rate MIT Licensed ~750ms Response

🔄 How It Works

The Reflexion Loop: Generate → Execute → Learn → Improve

📝
Your Task
"Calculate fibonacci(10)"
→
🤖
Generate
LLM writes code
→
đŸŗ
Execute
Run in Docker
→
✅
Success?
Check result
↩ī¸ If failed
🔍
Critique
Analyze error
↩ī¸ Retry (up to 3x)

✨ Features

What makes Agent Sandbox Runtime different

🔄

Self-Correction Loop

Automatically detects bugs, analyzes errors, and regenerates code until it works. Up to 3 retry attempts with learning.

🐝

Swarm Intelligence

5 specialist AI agents (Architect, Coder, Critic, Optimizer, Security) collaborate and vote on solutions.

đŸŗ

Docker Sandbox

Code runs in isolated containers with memory limits, no network, and automatic cleanup. Safe by default.

🔌

6 LLM Providers

Groq, OpenRouter, Anthropic, Google Gemini, OpenAI, and Ollama (local). Switch with one config change.

⚡

Fast Inference

~750ms average response time with Groq's LPU. 4x faster than GPT-4 Code Interpreter.

💰

Free to Run

Use Groq's free tier or run locally with Ollama. No cloud costs required.

🚀 Quick Start

Get running in under 2 minutes

One-line Docker run bash 
docker run -e GROQ_API_KEY=your_key ghcr.io/ixchio/agent-sandbox-runtime
Local installation bash 
# Clone and install
git clone https://github.com/ixchio/agent-sandbox-runtime.git
cd agent-sandbox-runtime
pip install -e .

# Configure
cp .env.example .env
# Edit .env and add GROQ_API_KEY

# Run
agent-sandbox run "Calculate fibonacci(10)"
Start API server bash 
# Start server
agent-sandbox serve

# POST a request
curl -X POST http://localhost:8000/execute \
  -H "Content-Type: application/json" \
  -d '{"task": "Check if 17 is prime"}'

đŸŽ¯ What It Can Solve

Benchmark-validated capabilities

Capability Example Status
Algorithm implementation Fibonacci, binary search, sorting ✅ 100%
Data parsing JSON extraction, CSV processing ✅ 100%
String manipulation Regex, formatting, validation ✅ 100%
Math operations Statistics, calculations ✅ 100%
Data structures Trees, graphs, lists ✅ 92%
Network/file access HTTP requests, file I/O ⚠ī¸ Sandboxed

📊 Benchmarks

Performance compared to alternatives

Tool Success Rate Avg Speed Self-Correct Sandbox Cost
Agent Sandbox 92% ⭐ 743ms ⚡ ✅ ✅ Free
GPT-4 Code Interpreter 87% 3.2s ✅ ✅ $0.03/1K
Claude 3.5 Sonnet 89% 2.1s ❌ ❌ $0.015/1K
Devin 85% 45s ✅ ✅ $500/mo