Dynamic Graph Compilation

[TimeLordRaps]

Checked

• I searched existing ideas and did not find a similar one
• I added a very descriptive title
• I've clearly described the feature request and motivation for it

Feature request

The ability to at run time dynamically insert/use newly generated agents in place for placeholder agents.
The ability to at run time dynamically insert/use newly generated tools in place of placeholder tools in newly generated agents..
The ability to at run time dynamically create/use newly generated vector databases as memory in newly generated agents with newly generated tools.

Motivation

I want a graph runtime where I have a lead programmer who is generally knowledgeable about the overview of the project who is capable of creating sub-agents who can then go and perform specific expert level tasks on a certain component of the software project.

Think like you have a lead programmer agent working at the javascript level who can spawn sub agents with expertise in langchain, who themselves could spawn experts in langgraph. At each level these agents would have different memories, that in theory should be constructable at runtime.

In this example the general graph structure is quite rigid, but I can imagine dynamic node/edge creation as an extension of this, however I'm not sure how that couldn't just be accomplished with placeholder nodes, or what it would provide in terms of additional functionality to my current proposal.

Proposal (If applicable)

DynamicAgent, DynamicTool, and DynamicMemory each which act as placeholders for to be defined behaviors. I think it would also make sense to create a toolkit which includes tools for generating predefined inputs to generate the dynamic objects so that it would be trivial to create a higher-level agent capable of creating sub-agents.

[xRME-18]

One approach to achieving this dynamic behavior is to leverage tools. Instead of relying on subGraph functionality, consider using tools that can compile and execute the desired agent.

However, one potential drawback is that you might not be able to take full advantage of all subGraph-related functionalities, as this approach disconnects the agents, it works for our use case. Although I haven't explored it extensively yet, we should be able to pass state even with this approach. I'm aware that state can be passed to tools.

Update: State can be passed to tools, but they can't edit state.