Handling goals and sub-goals as a decision tree in AutoGPT

[PriNova]

Duplicates

• [X] I have searched the existing issues

Summary 💡

To better handle goals and their sub-goals, I suggest using a nested to-do list that can read, add, and check items. This nested to-do list can then be integrated into the agent.json log or better as a seperated file.

I believe this will help the AI model to be consistent with the task and its subtasks to keep on track.

Examples 🌈

In my last test scenario to integrate NLP into an email summarizing process every morning, AutoGPT was stuck into reading <-> writing my email policy and procedure. I needed to hint to write a python script for gmail integration.

Another scenario was with planning, which resulted into 5 levels depth sub-tasks. AutoGPT lost the track because of a command loop. After the hint, the last fullfilled sub-goal was not deleted from the goals list or marked for solved, which leads to this loop.

Motivation 🔦

LLM is good at many tasks, but not on every task which needs to be strong deterministic. A nested to-do list is better written as a hardcoded system (for the moment) instead to rely on LLM (exclude planning). So I propose the feature of goal tracking as a simple nested list in python, which can be injected into a prompt and to keep track independently.

Later, this feature can be extended to an episodic memory, where the LLM commands are aligned and/or pre-fetched with learned task routines. This helps the LLM (and the user) to automate tasks already learned and successfully done.

This will extend to meta-programming, which leads to self-written plugins for AutoGPT.