Logo & Toolkit proposal

[adriangalilea]

Hello,

I've been thinking about what I denominated Squad(groups of agents) management frameworks, after dipping my toes in a few of them I think we needed something better. Instructor is so underrated for allowing LLM's to become foundations to build on top, after having played around with agency-swarm a bit, I really like what I see, good job.
(marvin.ai is really cool as an alternative to instructor too, check it out.)

I'd like to contribute:

Logo proposal

Ideation:
Took inspiration from hexagonal patterns because -> swarm -> bees -> hive ->hex.

Hexagonals are very geometric too, which fit the AI.

Also, self-similarity/fractals, because the future is dinamically spawning Squads that can spawn squads, and this will become self-similar and inter-connected.

So there's my logo proposal if you like it. Feel free to iterate, might seem too "snowflaky" but might work in the meantime.

On the technical side I have a few ideas to discuss and propose:

Toolkit

Some needs will be repetitive, so a framework such as this, should provide solutions for those common patterns. Let me give you an example:

The SearchWeb you proposed:

class SearchWeb(BaseTool):
    """Search the web with a search phrase and return the results."""

    phrase: str = Field(...,
                        description="The search phrase you want to use. Optimize the search phrase for an internet search engine.")

    # This code will be executed if the agent calls this tool
    def run(self):
        with DDGS() as ddgs:
            return str([r for r in ddgs.text(self.phrase, max_results=3)])

Is a very good start for the Toolkit

Let me share one, DownloadArticle:

from newspaper import Article

class DownloadArticle(BaseTool):
    """
    Download a news article and return its details.
    """

    url: str = Field(..., description="The URL of the news article to download.")

    def run(self):
        article = Article(self.url)
        print(self.url)
        try:
            article.download()
            article.parse()
            article_detail = {
                'success': True,
                'title': article.title,
                'authors': article.authors,
                'publish_date': article.publish_date,
                'content': article.text
            }
        except Exception as e:
            article_detail = {
                'success': False,
                'error': f"An error occurred: {str(e)}"
            }

        return {self.url: article_detail}

Many people use serper(bing with azure api too), for search, these options are fine, but duckduckgo works surpisingly good and is seamless. Also people use things like browserless for scrapping, and works great, but I found that Newspaper3k works wonders for the actual scrapping of articles and is another service you are freeing yourself from for the most part.

So the Toolkit would work like this:

from agency_swarm.toolkit import WebSearch, DownloadArticle

And just keep adding high value, vetted, hand-picked tools.

I have a few more ideas, but I think this is good enough for now.

What do you guys think?