AutoGPT

AutoGPT copied to clipboard

Duplicates

• [X] I have searched the existing issues

Summary 💡

Idea:

• Add a ai_settings directory to contain yaml files in the form of last_run_ai_settings.yaml. This way we could create and share interesting or successful "missions."
• Update scripts/main/parse_arguments so you can easily load any of the ...ai_settings.yaml files in the ai_settings directory.

Planning to add this to my local fork. Happy to do a PR if others like the idea :)

ps - awesome tool! thanks for sharing with the community!

Examples 🌈

No response

Motivation 🔦

No response

I think this is a great idea.

Regarding your second point there have already been multiple requests for this:

• https://github.com/Torantulino/Auto-GPT/pull/257#issuecomment-1497959394
• https://github.com/Torantulino/Auto-GPT/pull/257#issuecomment-1498239993

There's also a bug in determining where to save ai_settings.yaml #318

I think we need @Torantulino to weigh in on the desired functionality. Here's my proposal:

• [ ] Auto-GPT always saves the last setting it used to ai_settings.yaml.
• [ ] by default, Auto-GPT tries to load ai_settings.yaml and prompts the user if they want to use this setting again (this is the current behavior).
• [ ] An optional flag, e.g. --load_settings=./path/to/settings.yaml and Auto-GPT will try to load the specified file instead of ai_settings.yaml. A couple questions:
  • Since the user explicitly specified the file, should we skip asking the user for confirmation of the settings?
  • If the specified file can't be found, should we fall back to asking the user to enter the settings manually or simply error out?

I've got a pretty decent solution for this on my fork. https://github.com/gannonh/Auto-GPT

• Always saves settings to a unique file in a history directory
• Gives you the option at launch to ...

a - Load last used settings
b - Load another settings file
c - Create new settings

All options ask you to confirm your choice. Option b lets you select any file in the history dir. Happy to do a PR if there is interest. Thanks!

Screen recording of the feature.

New Recording - 4/7/2023, 7:34:43 PM

Went ahead and did a PR for this if interested: https://github.com/Torantulino/Auto-GPT/pull/453

This is kinda overlapping with the idea to come up with actual benchmarks for the project, including whole ai_settings.yaml files that may live inside the repository, also to be used for regression testing purposes: #1359

To be useful for benchmarking/regression testing, we would ideally be able to use these inside tests, too.

In other words:

• "Create a directory named Foo with a subdirectory named bar" -> should create Foo/bar
• "Create a file named test.txt" -> should create a file named test.txt
• "Rename test.txt to test.xxx" -> should rename the file

The point being, we can use an agent to come up with different mutations of the objective, to stress-test the engine - since the result always needs to be the same. And obviously, higher level features can then be built on top of these low level building blocks

Planning to add this to my local fork. Happy to do a PR if others like the idea :)

For the growth/success of the project, it will be even more important to share those objectives that didn't work out - so that the we can grow a suite of benchmarks to hopefully get rid of some of these issues over time.

This issue was closed automatically because it has been stale for 10 days with no activity.

FYI - implemented in Turbo: https://github.com/lc0rp/Auto-GPT-Turbo @gannonh, @dwrtz as personas / presets