A big, big restructure is coming!

[nicksolarsoul]

We are just 1 star away from 250 - what a week it has been since release!

One big learning to take away from this is that a multi-agentic approach is required in order to maximize the usefulness of the research. Already, I re-coded the project in CrewAI, but encountered too many limitations, so then I switched to AG2 (AutoGen) but encountered too many bugs.

One thing that became quite apparent is the state of multi-agent frameworks and their reliability. Often, I would encounter too many bugs, and the restrictions of the framework meant I could not accommodate some necessary fallbacks.

So, I've opted for a multi-agent approach with pure LangChain and using LangGraph, creating an orchestration method similar to here: https://github.com/langchain-ai/langgraph/blob/main/docs/docs/tutorials/multi_agent/hierarchical_agent_teams.ipynb

This will bring its own challenges, but due to plenty of ongoing LLM advancements like 2.5 Flash being both cheap and smart to make decisions, I feel I can quickly overcome them as they arise. LangChain/LangGraph comes with the maturity and flexibility to handle the situations when the framework doesn't deliver out-the-box.

With this in mind, I can create more roles and scale the platform in different ways, teaching it to be critical of its own research, fact-checking, verifying sources, and more through the use of agents with laser-focussed tasks on doing this.

Most of the research I ran checked out great, but there was one (in probably about 50) of the uses I ran that mis-attributed a source link. This is unacceptable to the perfectionist in me!

So hold off on the PRs - big change will be here in this week coming.

Of course, I want to keep it intuitive with ability to drop-in/drop-out multiple LLMs, and have it highly configurable to choose the depth of the research, enable/disable fact-checking, etc depending on preferences to either favor token efficiency or accuracy.

[nicksolarsoul]

About 90% of the way there now, a true agentic system is coming allowing you do drop in additional LLM models that autonomously collaborate with a "supervisor". Watch this space!

[fqolevelft]

Any updates Nick?

[nicksolarsoul]

Hi @fqolevelft I had to give a lot of focus on my paid client work but I'm back to it next week full time again. I am very nearly there but extensively testing (spending about $100 a day on testing the LLMs) and feel I have ironed out all the occasions where the LLMs make mistakes and pass invalid JSON or don't understand the context. I had severely underestimated agentic teams.

Then Google released A2A which I wanted to incorporate too - this thing is going to be a best, trust me! Another week or two.