Tagged
Agents, middleware, evaluation and releases across the LangChain stack. Curated and summarized from dozens of sources by AIBriefs.
LangChain's post argues human review is a strategic imperative for production agents, with three principles: expert judgment defines quality, agents must be evaluated on full traces including tool calls, and reviewer scores feed back into LLM judges and golden datasets. Flagged traces go to annotation queues in LangSmith, where a queue item can be a single run or a full multi-turn thread.
LangSmith Bring Your Own Cloud is GA on AWS and in beta on Azure and Google Cloud, keeping traces, evals, prompts, and agent deployments in the customer's own cloud account. LangChain runs the control plane (auth, billing metadata, hashed API keys) while the customer owns the data plane, and offers to run migrations from self-hosted.
Managed Deep Agents runs the open-source Deep Agents harness on LangSmith Cloud, now in public beta after a private beta that opened May 13, 2026. Agents are project folders tested with `mda dev` and shipped with `mda deploy`; durable memory, sandboxes and schedules are added as files rather than services.
LangChain CEO Harrison Chase argues the convergence on LLM-in-a-loop agent harnesses has bred a false sense that hard problems are solved, and that proactive, always-on agents bring new ones. He lists established fixes — thread summarization, offloading large tool results, skills, tool search, subagents, bash — and says no general solutions exist yet for the new set.
NemoClaw is a set of open blueprints, not a harness or runtime, that assembles an agent stack on top of a developer's chosen harness. It currently supports three: OpenClaw, Nous Research's Hermes Agent, and LangChain Deep Agents.
Managed Deep Agents v0.10 lets agents navigate, click, and fill forms on sites with no API, handing the browser window to the user for sign-ins so passwords stay out of model messages and traces. The release also adds async subagents that run long jobs in the background.
LangChain scaled from organic growth to five paid channels in six months, so it built an agent to reconcile cross-platform ad data and propose campaign changes. Calculations and safeguards live in code; the model handles judgment, with changes routed through human approval.
Agents CLI v1.8.0 ships eval generate, grade, compare, analyze and optimize commands to score agent traces and auto-tune prompts. The walkthrough's LLM-as-judge metric grades responses 1-5 on accuracy, relevance and clarity using gemini-3.8-flash.
LangChain's post breaks agent design into stacked loops: the core tool-calling agent, a verification loop with a grader that checks output against a rubric and returns feedback, and further levels instrumented with LangChain primitives. It cites Swyx's "loopcraft" piece and uses an internal docs agent that clones repos and opens pull requests as the running example.
On questions spanning multiple sources, an agent using LangChain's open-source OpenWiki wiki compiler as memory was ~5x more accurate than one using grep tools alone. Adding an evidence graph lifted accuracy another 27%, doubled citation consistency, and cut cost per correct answer 58%.
Google's developer blog walks through a working prototype, HappyDelivery, built on LangGraph, Gemini, FastAPI and Google Cloud across three decoupled layers: Ingress, Memory and Cognition. It includes deterministic financial tools and an automated verification suite, plus webhook handling for Meta's X-Hub-Signature-256 and Google's X-Goog-Signature.
Content Studio generates on-brand videos, graphics and decks from a single Slack prompt, with every asset written in code. LangChain spent four months teaching it design taste; each Slack thread gets its own sandbox that persists the project between turns.
LangChain built three internal agents — GTM Agent, OpenSWE for engineering, and Content Studio for marketing — all living in Slack, acting on behalf of users, with global and user-scoped memory. A large share of runs are scheduled rather than explicitly invoked.
LangChain's Deep Agents harness keeps searches small: glob matches file names, grep finds matching lines, and read_file reads a page at a time. Oversized tool results are offloaded to files, with the agent given only a path and preview.
Ramp spent months testing managed agent APIs, open-source frameworks, and third-party platforms before building Laminar, an internal agent harness that reached a first version in six weeks and now powers every agentic product at the company. LangChain's CEO called it part of a bigger shift.
LangChain's gateway sits between agents and model providers to enforce spend caps, rate limits, model fallbacks and PII/secret redaction in one place. It spans closed and open-weight models to avoid provider lock-in.
Alter gives each agent its own scoped identity and keys instead of one shared over-scoped credential, with default-deny policy checks on every request. One SDK call handles storage, refresh, policy and audit for 140 providers, with LangChain, MCP and AWS integrations.
Agents CLI 1.5.0 introduces an Extension System that overrides or adds commands without forking the tool; 1.8.0 ships a first-party LangChain extension replacing framework-coupled commands. Teams keep the unified deploy, eval, and observability workflow without adopting ADK.
UK digital health firm Numan replaced single-turn static-fixture evals with per-category, simulated-conversation evals for Nu, its AI health coach, after averaging one score hid failures. One low-score cluster traced to the rubric, not the agent: it demanded empathy, exploration and advice within two turns.