Auto Pilot is a method aimed at analyzing the repetitive tasks of AI agents and executing these tasks automatically, much like a human would. Currently, there is not much focus on metrics such as performance for AI agents; however, particularly in conversable agents, a tendency towards “overthinking” has been observed. As a result, even in very simple tasks, an excessive amount of “generation” occurs, leading to a decrease in performance. Auto Pilot seeks to store frequently repeated tasks in memory so that when a similar or identical task is encountered again, it can swiftly apply the steps learned from previous processes without further deliberation.
You are currently viewing the proof-of-concept (POC) repository for this idea. The implementation is still quite basic and primitive, but it is expected that with further development, it will become a general method applicable to all kinds of agentic AI frameworks. At present, tests are being conducted using the Autogen framework.
We anticipate that, in the future, this method will serve a role similar to RAG (Retrieval-Augmented Generation) for large language models (LLMs) or to “functions” in programming languages. While agents are already impressive without Auto Pilot, performance evaluations and process optimizations will become more pressing over time. This is where Auto Pilot and similar approaches are expected to come into play.
@software{ AutoPilot,
author = {Bunyamin Ergen},
title = {{AutoPilot}},
year = {2025},
month = {01},
url = {https://github.com/bunyaminergen/AutoPilot},
version = {v0.1.0},
}