SuperHarness - AI Agent Development Framework

Overview

SuperHarness is a complete AI Agent development framework built with Rust + Python, benchmarked against Claude Code and LangChain.

• Python SDK: Clean API for building AI applications
• Rust CLI/TUI: High-performance terminal interface
• Multi-LLM Support: Anthropic, OpenAI, Gemini, custom endpoints
• Complete Toolchain: File operations, code editing, shell execution

Six-Layer Architecture

┌─────────────────────────────────────────────┐
│ Layer 5: Application Layer (CLI/TUI)        │  ← User Interaction
├─────────────────────────────────────────────┤
│ Layer 4: Agent Runtime                      │  ← Agent Runtime
├─────────────────────────────────────────────┤
│ Layer 3: Tool System                        │  ← Tool System
├─────────────────────────────────────────────┤
│ Layer 2: Session & Memory                   │  ← Session & Memory
├─────────────────────────────────────────────┤
│ Layer 1: Core Services                      │  ← Core Services
├─────────────────────────────────────────────┤
│ Layer 0: Security Foundation                │  ← Security Foundation
└─────────────────────────────────────────────┘

Python SDK

from superharness import Agent
 
# Create Agent
agent = Agent(
    api_key="your-api-key",
    provider="anthropic",  # or "openai", "gemini", "custom"
    model="claude-sonnet-4-6"
)
 
# Simple conversation
response = agent.run("Hello, how are you?")
print(response)
 
# Streaming output
for chunk in agent.run_stream("Tell me a story"):
    print(chunk, end="", flush=True)
 
# Register tool
@agent.tool
def calculate(expression: str) -> float:
    """Evaluate a mathematical expression."""
    return eval(expression)
 
# Use tool
response = agent.run("What is 123 * 456?")

CLI / TUI

# Enter TUI directly
superharness
 
# Show help
superharness --help
 
# Show version
superharness --version

Development Progress

Phase 0: SDK/TUI Basic Features ✅ Complete

Phase 1: Production Features ⏳ In Progress

Design Philosophy

Core Principles

1. No MVP/Demo - Build complete product directly
2. Benchmark Claude Code/LangChain - Match or exceed
3. Precise Tasks - Clear completion criteria and acceptance conditions
4. Real Validation - Validate with real user workflows

Transparency Three Principles

• No Hidden Behavior - All operations visible
• State Visualization - Real-time progress display
• Explainable Decisions - Explain tool selection and execution

Current Status

Phase 0 complete, 192 tests passed, Phase 1 in progress.