rfc: Per-agent context splitting

[pkarw]

Right now we've got a shared context per-team. All agents have access to all information from the execution context. This is great, and simplifies a lot - but we'll shortly deal with:

• context windows issues per LLM,
• discrepancies in the case that agents use different LLMs (llama_vision vs gpt-4 vs mistral etc) - idk for sure, but it could be a real mess when multiple LLMs will base on the results generated by other ones

[pkarw]

This is a rfc - and let's discuss if this makes sense. Will make the #1 a little bit more complex, requiring some kind of agent tagging but it's not really an issue

[grabbou]

BTW, agents have their own context, they only inherit "current state" of workflow to get started. It gets longer of course with more tasks, but it is for context. We could decide if it is needed, or we change it to "task" and "context" (all necessary information to perform task).

Once agent is done processing, it returns its answer, so agents have their per-agent context.

[pkarw]

I'm closing it for now, I think the current way of context switching is exactly what I had in mind creating this issue before getting into the game deep enough 😅