Thanks Harrison. LangGraph (eg graph theory + Networkx) is the correct implementation of multi-agent frameworks, though it is looking further into, and anticipating a future, then where most GPT/agent deployments are at.
And while structured output and tool calling are good, from client feedback, I'm seeing more of a need for different types of composable agents other then the default ReAct, which has distinct limitations and performs poorly in many scenarios. Reflection/Reflextion are really good, REWOO or Plan/Execute as well.
totally agree. we've opted for keeping langgraph very low level and not adding these higher level abstractions. we do have examples for them in the notebooks, but havent moved them into the core library. maybe at some point (if things stabilize) we will. I would argue the react architecture is the only stable one at the moment. planning and reflection are GREAT techniques to bring into your custom agent, but i dont think theres a great generic implementation of them yet
CIO's or similar. We've been talking to a number of directors at consulting firms who say they are getting calls daily from CIO's who share with them the enormous pressure that they are under, receiving calls from the C-suite and even board of directors asking "what are we doing about generative AI?"
I think you're correct in that its overused, but the jump from using a spreadsheet and programing even the simplest macros, is a big one for a large portion of the those use Excel. Much less using a RDBMS or programming in Python. Even trying the macro recorder makes most users assume 'that's not meant for me'. If the requirement of the job include building a model that can be used and manipulated by anyone than Excel is the only tool for the job, anything else requires more training.
It's a custom developed planning/forecasting tool. It's a P x Q thing, but the P and Q are variable as to the type and mix of the Ps and Qs and need to be combined/sliced into many different ways for analysis/planning/forecasting. The team actually developing it has been very challenged.
You're not really challenged until the growth rate of their Excel monster outpaces the rate at which you automate the existing stuff... thankfully Excel has a 1 million row limit, and it always takes them time to figure out how to shard an Excel database :P
Well in this case it's actually mine and the rest of the finance group's Excel monsters (plural) that are being automated. The devs are doing yeoman's work.
And while structured output and tool calling are good, from client feedback, I'm seeing more of a need for different types of composable agents other then the default ReAct, which has distinct limitations and performs poorly in many scenarios. Reflection/Reflextion are really good, REWOO or Plan/Execute as well.
Different agents for different situations...