LangChain

LangChain rebuilt its MCP support around the July stateless spec: MCP moves from the separate langchain-mcp-adapters install into langchain.mcp in the main package (langchain[mcp]>=1.4.0, beta), MultiServerMCPClient collapses into one MCPAdapter rebuilt on FastMCP, and per-connection negotiation covers servers on both protocol generations. It adds elicitation via LangGraph interrupts and client-side tool-list caching. Python ships first, TypeScript to follow. ↗

LangChain packages the internal process it uses to build synthetic agent environments and tasks into an updated eval-engineering skill. Environments define where the agent runs; tasks pair an input with an environment and a test script whose output is scored; together they form a benchmark. Teams can now own the same flow LangChain uses internally for evaluating agents before shipping. ↗

LangSmith Engine now detects more than 2× the impactful issues on IssueBench, writes fixes 25% better on Terminal-Bench, and ships with self-hosted support, Slack alerts, automatic Linear ticket creation, a reduced-analysis mode, and auto-closing of stale issues. Since its May launch, Engine has scanned over 60 million traces and surfaced more than 20,000 issues across customers' agents. Available on LangSmith Plus and Enterprise, in SaaS and self-hosted. ↗

OpenWiki's new memory layer treats forgetting as essential: every claim in a wiki page is linked to a versioned evidence reference in the source code, and on each update the runtime deterministically checks whether the evidence still matches. Stale claims are either refreshed or rewritten; unresolved ones stay flagged. Replays on a 2000-claim benchmark show supported claims rise from 92.9% to 97.8%, and hallucinated claims drop from 0.7% to 0%. ↗

Preview Builds spin up temporary, production-like environments from PR branches, each linked to the parent deployment, so teams can test agent changes before merging. New commits refresh the preview, collaborators can interact with the running agent without cloning the repo, and idle TTL plus max-concurrency caps keep resource use in check. Public beta for GitHub-integrated deployments on LangSmith Cloud. ↗

LangSmith Tuned Evaluators are managed, versioned evaluators that auto-attach quality feedback to production traces and threads. The launch evaluators — Perceived Error — flag threads where the agent made a mistake, misunderstood a request, or derailed; the specialized post-trained judge reportedly outperformed every frontier model in LangChain's benchmark at up to 98% lower cost in some workloads. Available now for Plus and Cloud Enterprise plans in the US. ↗

AgentCore Payments is middleware that lets LangChain agents reach paid APIs and services while enforcing deterministic, session-level spending limits. When a tool calls a paid endpoint and gets HTTP 402, the middleware checks the charge against the session budget, signs the payment if affordable, and retries the request with payment proof — with the budget enforced by infrastructure, not the model prompt. ↗