LangChain AI Updates: October 15, 2026

1. LangGraph Postgres Checkpointer 3.2.0 Adds Checkpoint Pruning

LangChain. langgraph-checkpoint-postgres 3.2.0 implements prune and aprune on PostgresSaver and AsyncPostgresSaver, which previously raised NotImplementedError and left long-lived threads with no supported way to drop old state. The delete strategy removes every checkpoint, pending write, and blob for the given threads, while keep_latest keeps only the latest checkpoint per namespace plus the writes and blobs it references; finished subgraph namespaces are removed along with their root. Each thread is pruned in a single REPEATABLE READ transaction, and the PR author reports about 100 to 200 ms for a 6,000-checkpoint thread locally. keep_latest refuses to run on threads using a DeltaChannel, and Postgres will not shrink table files without VACUUM FULL or pg_repack. Source

2. LangChain Details How Deep Agents Manages Context With Middleware

LangChain. A LangChain engineering post walks through the middleware Deep Agents uses to keep long-running agents from degrading as context grows. FilesystemMiddleware writes tool results over about 20,000 tokens to /large_tool_results/{tool_call_id} and shows the model a preview and path, while SummarizationMiddleware condenses older messages once usage reaches 85% of the context window and saves the summary to /conversation_history/{session_id}.md so it can be recovered. SkillsMiddleware loads only skill headers at startup, subagents keep noisy work in a separate context, and MemoryMiddleware keeps long-term memory in a markdown file such as /memories/AGENTS.md that stays portable across model providers. The post describes existing capabilities and does not announce a new release. Source

3. LangChain Publishes Agent Evals FAQ Built Around LangSmith Workflows

LangChain. LangChain published guidance on common agent evaluation questions, recommending teams start with 10 to 20 examples per behavior drawn from real traces, record a baseline, and feed failures caught by online evaluators back into offline datasets. It advises code checks for deterministic properties, LLM judges limited to narrow yes/no or small categorical scales and calibrated against human labels, and scoring multi-step agents on tool-call trajectories and final state, not just the final answer. The post cites a Jev-as-a-Judge run that matched human pass/fail labels on every decision for $0.34 total versus $28.17 for Claude Sonnet 4.6, and maps each practice to LangSmith features such as annotation queues, online evaluators with sampling, and pytest and Vitest integrations. Source