Aww: automatic self-improvement of prompts

I implemented self-improvement of prompts in Augmented Awareness. It uses two LLM agents, one acts as the generator and a second one as the critique. The critique receives the prompt, the input and the output of the generator agent, and generates an improved prompt.

It is still an open-loop control for now, meaning the critique doesn’t receive human feedback.

The code looks like the following:

def rewrite_prompt(...):

    # ...

    critique_model = make_model(...)
    critique_agent = Agent(model=critique_model)

    @critique_agent.system_prompt
    def critique():
        return textwrap.dedent("""
        You are an expert at writing LLM prompts. You will receive:
        1) the prompt
        2) the output
        3) a series of input messages
        Your job is to write a revised prompt that is more performant.
        Write only the revised prompt in full, in markdown format.
        """)

    prompt = prompt_file.read_text()

    llm_model = make_model(...)
    gen_agent = Agent(model=llm_model, system_prompt=prompt)

    async def do_critique():
        gen_result = await gen_agent.run(user_prompt=content)
        gen_output = gen_result.output
        critique_result = await critique_agent.run(user_prompt=[prompt, gen_output] + content)
        return critique_result.output

    revised_prompt = asyncio.run(do_critique())
    prompt_file.write_text(revised_prompt)

I believe the principle of AI productivity is that AI systems should be able to improve themselves. That’s how exponential progress can be achieved.

Balancing errors and improvements is the challenging part. That's why it remains a human-in-the-loop process. For Aww, this involves manually analyzing the results and making git commits of the revised prompts.

Aww is a recursive system that uses four different agents to create daily, weekly, monthly, and yearly retrospectives, which build on each other. Future updates will aim to make self-improvement and self-adaptation automatic and cover multiple levels.