Adding Agents

[danielmiessler]

One thing we're really looking to do is add agent functionality to Fabric.

Specifically, I want to have endpoints that point to things like plan_trip, for example, which actually runs a CrewAI team to do that task.

We would have all the required keys in the ~/.config/fabric/.env file so that it just works.

Thoughts?

[dheerapat]

something like #44 ?

[danielmiessler]

No, more like CrewAI, or Langraph, or Autogen.

[chrissonntag]

I've created a generic AutoGen setup today: https://github.com/chrissonntag/yaas

It's still in early days, but I'm thinking it could potentially be integrated into fabric. With this setup, we can configure workflows using simple MD files, making everything super easy to manage. The key to getting consistent results is giving clear and quality instructions. So if we manage to do so via appropriate patterns, surplus implement the one or another "skill", we should get it going 😁

[johnconnor-sec]

Any updates on integration with crewAI?

[johnconnor-sec]

I see praisonAI demos often using a SERP API key. How does fabric achieve web retrieval? Is there a way to define what tools are available to praisonAI and thus fabric?

[johnconnor-sec]

Does this mean my praison agents will be able to be imported?? 🤞 I'm new to python but this is so cool i have to learn it

[johnconnor-sec]

So if I understand the stack correctly - I would use LangGraph -> CrewAI -> PraisonAI -> fabric as the dev stack. Could I train models using this stack? Is it possible for a small team to specialize models for specific tasks to accomplish specific goals, possibly by renting GPUs?

In theory, could I train many smaller models on Boolean questions then train more models with slightly more expertise on the same subject to work as managers. If so, I could also implement rag for each, possibly with regex. Thinking if an army of "Russian dolls" models trained for specific tasks, possibly led by hierarchical "need to know" managers.

[johnconnor-sec]

Can fabric be used for training ?

[johnconnor-sec]

Whoa 😳 I think I'm getting a better idea of what I'm working with, i.e. --agents 🤯

[johnconnor-sec]

I think adding doraindarko/maestro would be a wonderful idea. Is there a way to include my own agents in fabric? I now have agents to do tasks and would like fabric to have the ability to call them like plan_trip.py. If I place an agent file in /agents, will that work?

[geektx]

Does this mean my praison agents will be able to be imported?? 🤞 I'm new to python but this is so cool i have to learn it

Have you discovered a method to bring them in? I have created a pattern that creates agents.yaml playbooks that I really want to pipe to --agents.

[johnconnor-sec]

I believe I have but have yet to test it

Might try tonight

[nic0711]

How about integrating something like this?

mendableai/firecrawl

[prabhic]

Today have tried
AGENT TEST
echo "generate qt c++ code to show 100 qml delegates horizontally. qml delegate contains image widget, and passe to c++ code. widget code should be in purely c++ without using any qt default listview framework. and should implement using qt quick framework without default listview. it should be optimized for memory and performance" | fabric --agents

OUTPUT
Firstly I am Impressed to see that agents are created automatically. and the code is generated successfully. Created following Agents. [DEBUG]: == Working Agent: QML Designer, [DEBUG]: == Working Agent: C++ Developer, [DEBUG]: == Working Agent: Performance Optimizer. and code is generated.

FEEDBACK
Understandably, generated code have deviation from my expectation. Mostly it can be because of the clarity in my input prompt.

SUGGESTION
However If we are able to iterate on agent description and role, and task, on command line itself, it will have huge uplift.

Nice to see and experiment on fabric, great work.

[ajoslin103]

@prabhic are you using the Python or GO version of Fabric ?