A minimal harness for agents

harness.py

import subprocess
import re

# 1. Initialize Memory
messages = [
    {"role": "system", "content": "You are an agent. Write code in ```python filename.py format. I will run it and tell you the result."}
]
max_iterations = 3
current_iteration = 0

while current_iteration < max_iterations:
    # 2. Call the Model (e.g., OpenAI, Anthropic, or local model API)
    response = call_llm(messages) 
    messages.append({"role": "assistant", "content": response})
    
    # 3. Parse and Write File
    code_blocks = extract_code_and_filenames(response) # simple regex
    for filename, code in code_blocks:
        with open(filename, "w") as f:
            f.write(code)
            
    # 4. Execute and Capture Feedback (The Deterministic Loop)
    result = subprocess.run(["python", "test_target.py"], capture_output=True, text=True)
    
    if result.returncode == 0:
        print("Success! Tests passed.")
        break # Exit the loop, task complete
    else:
        # 5. Feed the error back into the model's memory
        error_message = f"Execution failed:\n{result.stderr}\nPlease fix the code."
        messages.append({"role": "user", "content": error_message})
        
    current_iteration += 1

if current_iteration == max_iterations:
    print("Agent failed to solve the task within iteration limit.")