Keep pulling the thread on Erick Friis.
Tool calling is currently the most important LLM feature for LangChain and LangGraph users because it provides a crucial structured interface between code and the models.
The biggest hurdle to productionizing LLM-based applications is reliability, stemming from the non-deterministic nature of model outputs.
LangChain's main focus has shifted to LangGraph, a tool for orchestrating AI agents as state machines.
The simple ReAct agent loop performs poorly as the number of tools increases, causing the LLM to call the wrong tool or use incorrect parameters.
A recommended best practice for building reliable AI agents is to limit an LLM's access to approximately five tools at any given step in a workflow.
Agentic flows built with LangGraph can enable the use of smaller LLMs, which provides cost savings and allows for execution on less powerful hardware.
LangGraph includes a default recursion limit to prevent agents from entering an endless loop, analogous to a function call stack limit.
The 'human in the loop' interruption capability is now a first-party, natively supported concept within LangGraph.
LangChain's experience with agent memory suggests a 'less is more' approach is most effective, favoring simple key-value storage over complex features like automatic conversation history trimming.
LangChain's development philosophy has evolved to be more 'bare bones', favoring raw Python and JavaScript to reduce opaque abstractions and make the framework more intuitive for developers.
The tool-calling performance of Meta's open-source Llama models varies significantly across different hosting providers due to inconsistencies in how they implement tool call parsing.
LangChain offers a commercial product called LangGraph Platform for hosting agent applications, analogous to the relationship between the Next.js framework and the Vercel hosting platform.