search

TransformerOptimus / SuperAGI

<⚡️> SuperAGI - A dev-first open source autonomous AI agent framework. Enabling developers to build, manage & run useful autonomous agents quickly and reliably.
https://superagi.com/
MIT License
14.93k stars 1.78k forks source link

Agent should be able to create a tool autonomously #61

Open Prasaanth-Vallabhan opened 1 year ago

Prasaanth-Vallabhan commented 1 year ago

Agent should be able to programatically create a tool, load it dynamically and use on the fly.

Cptsnowcrasher commented 1 year ago

This can be packaged as a toolkit which could constitute of access to tools apis, validation of the tool , iterate and improve it

georganix commented 1 year ago

love the idea! let us know when ready!

sdrakulich commented 1 year ago

https://github.com/Nicolas-Gatien/MUSE worth checking out

Sanchay-T commented 11 months ago

I have been giving some thought to our current approach towards tool development and I'd like to suggest a different strategy that could potentially make our process more efficient and cost-effective.

Instead of creating a wide variety of specialized tools, we could look into using a restricted execution environment, such as 'RestrictedPython'. This environment would come pre-installed with the necessary libraries required by our tools, and provide a secure and controlled space to execute our code, similar to the function of a code interpreter.

One of the key features we could implement within this environment is a filter mechanism. This mechanism would act as a safeguard against the download and usage of potentially harmful libraries, ensuring the safety and integrity of our environment. Although this might limit the use of unrestricted 'pip install', I believe we can address this by implementing appropriate barriers or filters.

To help guide the construction and use of our tools, I suggest maintaining an extensive, step-by-step documentation. This documentation can serve as an initial resource for our universal agent. As we develop new tools or update existing ones, their respective documentations can be added to this resource.

Additionally, to assist the universal agent in building a particular tool, we could employ techniques like Maximal Marginal Relevance (MMR) to search for the most relevant documentation. This approach will allow the agent to find the most pertinent information effectively.

Looking forward, I anticipate advancements in AI, such as GPT-5 or its successors, which will likely be trained on comprehensive internet data. These advancements could naturally solve our current challenges. However, until such technology is available, the above-proposed solution could serve as an effective, safe, and cost-efficient approach.