LangChain AI Updates: May 6, 2026

1. Harrison Chase Argues Agent Observability Is Incomplete Without Feedback

LangChain. Harrison Chase published a piece arguing that traces alone “tell you what happened” but not whether it was good — and that agent observability platforms must combine trace storage, attached feedback storage, and automated feedback generation (rules, LLM-as-judge) to drive learning at the model, harness, and context layers. The post reframes feedback as the raw material for systematic improvement rather than a satisfaction rating tacked on at the end. Source

2. LangGraph SDK 0.3.14 Adds return_minimal for Threads Update

LangChain. The LangGraph Python SDK released 0.3.14, adding a return_minimal parameter to threads update that lets clients skip large response payloads when only acknowledgement is needed. The release also bundles alpha pre-releases of langgraph, langgraph-checkpoint, and langgraph-checkpoint-postgres for users tracking the 1.2 line. Source

3. LangGraph Checkpoint SQLite 3.1.0a1 Reworks Delta Cadence

LangChain. A new alpha of langgraph-checkpoint-sqlite (3.1.0a1) ships a streaming-walk approach for delta channel history retrieval and promotes get_writes_history to a public saver API, aligning the SQLite backend with the broader delta-cadence rework happening across LangGraph 1.2. The change targets large-graph performance where prior implementations materialized too much history per read. Source