langchain langchain==1.5.0

LangChain 1.5.0 adds two pieces of agent middleware, code that sits around an agent's model and tool calls, and both depend on a new kind of component LangChain calls a decision model. A decision model is a classifier: given some input, such as the user's latest message or a pending tool call, it answers a fixed question, such as which of several options fits or how likely something is to be true, and returns probabilities instead of generated prose. The companion langchain-core 1.8.0 release adds a shared interface for these models, so the middleware can work with any classifier that implements it.

The model router asks the decision model a multiple-choice question about the user's latest message and sends the rest of that agent run to whichever model it picks. Auto mode screens tool calls before they execute: for tools the developer lists, it asks the decision model how likely the call is to be risky, and at 0.5 or above it refuses to run the tool and hands the agent an error instead. Both were ported from the integration package for TypeSafe AI, a lab that builds models for making decisions inside software, whose first LangChain release came out a week earlier. TypeSafe's classifier is the only implementation the pull requests name, so this release supplies the slots more than a choice of models to fill them.

The router now prints a beta warning because its interface may still change. The release also ends support for Python 3.10, which has reached end of life; projects still on 3.10 will keep installing the 1.4 series rather than break.