I built a Deep Research Agent with Jev as the review layer! Nemotron does the research, Tavily for the web search, and Jev decides whether the research is actually good enough to return. The agent starts with a question, searches across multiple sources, and builds an initial answer from what it finds. But instead of immediately sending that answer back, it passes the research to Jev. Jev looks at the result as a decision problem: is the answer sufficiently supported, does it cover the original question, and is there enough evidence to stop researching? If the answer is no, the agent gets another research pass. So the workflow becomes: Research → Draft → Jev Review → Research again if needed → Final answer That separation is useful because the model doing the research doesn’t also get the final say on whether its own work is complete. For the stack, I used Nemotron through Nebius Token Factory for the main model, Tavily for the web search, LangChain to run the research loop, and LangSmith to trace what happens across each step. I built this while trying out the Nebius AI Builder Program. What made sense for this project is that the program is built around this kind of stack: open models through Token Factory, retrieval with Tavily, orchestration with LangChain, observability with LangSmith, and other components that can be swapped independently instead of locking the whole agent into one setup. The full project is open source. I’ve shared the GitHub repo in the comments!
LangChain on X
Builders who credit LangChain in working agents and praise its harness writeups lead the talk; a smaller, named group of critics calls it over-abstracted or past it. People ask whether it is still the way to go, how its numbers are measured, and how agent auth works.
First written 27 Oct · what changed · from 92 posts by 78 people
Tracking posts that say “langchain”
Track this for your marketPosts per day, last 30 days
as of Nov 3, 09:00 UTCWhat changed
as of 27 Oct · checked once a monthFirst written 27 Oct. When what people say changes, it shows here.
Tone· 43 posts from 40 people· written 27 Oct
How do people feel about LangChain?
Fans lead: 18 of the 40 people in this section credit LangChain, LangSmith or LangGraph in a build that works, and the harness and writeups draw a further 11. A named minority of 8 call it past it or beside the point, and 3 are openly unsure it still fits.
What fans report is concrete: middleware for summarisation and human-in-the-loop, a first RAG chatbot over YouTube transcripts, deterministic scripts over LLM skills, and a stack that keeps LangChain, DeepAgents and LangGraph for what each does. LangSmith is named for dumping OpenTelemetry spans instead of rebuilding a custom tracer.
Praise for the harness is about defaults and details, not models: an agent that queries Salesforce on the company account but sends mail only as the user, code running in a sandbox behind a security proxy, memory split into global and personal layers, agents living in Slack. The writeups draw the strongest words — "the best field notes in this space" — and one post credits a jump from 53% to 67% on a coding benchmark to changing only the harness.
The critics are a minority: over-designed abstractions, "nobody knew what a harness was", LangChain gone cold, and the wrong first thing to learn — one advises learning LLMs, RAG and evals before it, since jumping in first makes nothing make sense.
- Fans and happy users19
- Praises the harness and writeups13
- Past it or beside the point8
- Undecided if it still fits3
See 5 of 43 postsHide the posts
最近 LangChain 分享了内部销售、研发和内容团队的 Agent Harness 架构，里面有不少值得借鉴的工程细节。 他们发现，最棘手的问题之一是「Agent 到底以谁的身份干活」。 比如销售 Agent 查询 Salesforce 可以用公司统一账号，但发邮件必须使用用户自己的授权，而且只能在私聊中操作。 Coding Agent 也一样。帮你提交 PR 时，用你的 GitHub 身份；在后台自动审查代码时，则切换成独立的 GitHub App。 内容 Agent 发公司 Twitter，用的是公司账号，但只有白名单用户能真正发帖。 除此之外，还有几项关键设计： 1. Agent 直接接入 Slack，不用员工换平台；需要时还能从工讨论里「拆」出独立线程 2. Memory 分全局和个人两层，由 Agent 自己决定一条事实记到哪一层 3. 代码在沙箱里执行，经安全代理访问外部服务 4. 大量任务是定时或事件触发的，不一定等人开口。比如新 PR 自动 Review，或者你跟它说一声，它自己挂上每天销售简报 Demo 阶段跑得通的 Agent，一进公司就会撞上身份和权限边界。这些细节平时讨论不多，但真落地一个都绕不过去。 这些细节平时讨论得不算多，但真到落地时，一个都绕不过去。推荐做企业 Agent 的朋友读读。 原文：link
LangChain jumping from 53% to 67% by changing only the harness is the useful example. Most people still treat the model as the only variable. Have you gone through any of the steps, or is this still the announcement?
Probability of getting rejected for not knowing LangChain, RAG, and AI agents: 20% Probability of getting rejected for not understanding concurrency, memory management, caching, and database indexing: 99%
I was thinking about this for a while, but I couldn't keep it to myself: Do you guys remember LangChain? What the heck happened to them. Last year I was going to propose this for use at work, but now I don't even hear about it. This space moves SO quick 🥲
5 of the 43 posts behind this answer, one per person. The other 38, and every new one as it’s posted, come with tracking. Get early access
Pain points· 21 posts from 19 people· written 27 Oct
What do people complain about with LangChain?
No single complaint leads. The biggest is that building with it is a slog (4 people), and behind it three are level at 3 people each: auth and governance at scale, retrieval defaults costing accuracy, and lock-in with dependencies nobody signed up for.
For the largest group the friction is hands-on: errors while learning tool calling, a credit requirement blocking langchain-typesafe, and a bug fix assigned back to the reporter with the PR left "waiting on review for 6 weeks".
Auth is called "the annoying part", and the worry is scale: teams stumble on the governance layer, and running hundreds of agents across thousands of concurrent users is where it gets hard. On retrieval, the complaint is that defaults turn memory to noise — no recency decay or frequency weighting, and fusion as the shipped default costs accuracy on some queries; reliable recall is said to need provenance, timestamps and expiry.
Lock-in shows up as platform teams mandating an SDK that boxes LangChain stacks out of deployment and eval pipelines, and as hidden pulls: a locally run agent still fetching its prompt from the Hub. Smaller groups name docs that go stale (one credits the v1 with fixing previously poor docs), breakage with no error — a langchain-core 1.7 streaming tool-call change that leaves dashboards quietly reading zero — and traces keeping full tool responses, with one person asking whether the tracer masks keys or masking has to happen upstream.
- A slog to work with4
- Auth and governance get hard at scale3
- Retrieval defaults hurt accuracy3
- Lock-in and hidden dependencies3
- Bad documentation2
- Quiet breakage on upgrades2
See 5 of 21 postsHide the posts
filed a bug on vercel's eve harness. they picked it up and shipped the fix in 3 days filed one on langchain. they assigned the fix to me, i wrote it, and the pr's been waiting on review for 6 weeks so i think we know which agent harness is gonna win
retrieval without recency decay or frequency weighting is why most agent memory setups turn into noise after a week. context engineering is way more about what you throw away than what you pull in.
Framework lock-in is where enterprise AI agent projects go to die. When platform teams mandate an opinionated SDK, teams using LangChain or custom stacks get boxed out of deployment and eval pipelines. Today Google opened up the Agents CLI Extension System: bring any framework, keep unified Cloud Run deployment.
Sounds promising, but most teams stumble on the governance layer once they scale
不报错的升级最危险：LangChain core 1.7 和 ADK 2.13 都改了格式，程序照跑，监控会安静地失明。 langchain-core 1.7.0 把流式工具调用参数改成 tool-call-delta 分片下发。 以前是 block-delta 快照。 官方把它列为破坏性变更。 Google ADK Python 2.13.0 把默认遥测换成 OpenTelemetry GenAI semconv 1.44。 call_llm 这个 span 直接移除了。 两个版本只隔一天发布：LangChain 10-23，ADK 10-24。 LangChain 这条，影响的是从流里拼工具参数的代码。 ADK 这条，影响的是按 span 名字统计调用次数和耗时的查询。 按 call_llm 去查的面板，查到的就是空。 如果你的成本看板或告警，是靠解析这些流和 span 算出来的，升级之后不会有任何报错。 只是数字不再更新。 看板上的 0，看起来和“今天很安静”一模一样。 两个都有回退办法： LangChain 锁 langchain-core<1.7。 ADK 设 ADK_TELEMETRY_SCHEMA_VERSION_OPT_IN=otel_semconv_1_36。 报错的升级，当天就会被修。 不报错的那种，会一直错到有人去看数字。 升级后的第一项检查，不是程序能不能跑，是看板上的数字还在不在跳。
5 of the 21 posts behind this answer, one per person. The other 16, and every new one as it’s posted, come with tracking. Get early access
Alternatives· 14 posts· written 27 Oct
What are the alternatives to LangChain?
CrewAI is the biggest alternative named, at 3 people, on where each fits: CrewAI for multi-agent systems, LangChain and LangGraph for RAG and agentic workflows. Behind it, Claude Code, a plain while loop with a strict tool schema, and Google's Agents CLI Extension System are level at 2 people each.
The CrewAI comparison is a slotting question more than a rejection: one person measuring their agent picked LangChain + LangGraph + Mem0 because "CrewAI was more about multi-agent systems, LlamaIndex is for RAG agents", another rates LangChain, LangGraph and LangSmith good for both agentic and RAG work while allowing CrewAI is good too, and a third keeps going "back and forth" on CrewAI or LangChain for multi-agent.
Claude Code appears as the thing to reach for before building: agents made over weeks are called less useful than the off-the-shelf tool, and one person says they are not on LangChain, with Claude driving their own code instead. The while-loop argument is that frameworks churn every three months, so root yourself in loop, tools and state; one post notes 0 of 11 teams using an agent framework.
Google's Agents CLI is read as an opening rather than a rival — a first-party LangChain template, deployment kept with Cloud Run, and an end to being boxed out by a mandated SDK. The rest are single mentions: LlamaIndex, Jev ("shipped before the right primitives existed"), Eino for Go teams, Google ADK as the host LangChain plugs into, and Managed Agents ranked beside Claude's.
- CrewAI3
- Claude Code2
- Just a while loop2
- Google Agents CLI2
- LlamaIndex1
- Jev1
See 5 of 14 postsHide the posts
these are current measurements of my ai agent. tomorrow or after-morrow, hopefully with migrated project, i'd show you a comparison and that would be v1. i chose Langchain + LangGraph + Mem0, because CrewAI was more about multi-agent systems, LlamaIndex is for RAG agents.
crewai or langchain for multi agent stuff? i keep going back and forth
真的觉得现在很多人学 AI 开发，把顺序搞反了。 实际问题还没解决，就急着学 LangChain、搭 MCP、设计 Multi-Agent。 最后花几个星期造出来的 Agent，根本不如现成的 Claude Code 好用。 如果你的目标是解决业务问题，先把现有工具的能力用到极致。 等它真的满足不了需求，再考虑自己造轮子。
0 of 11 using an agent framework is a rough sentence for everyone who spent last year writing LangChain tutorials. Turns out the secret sauce is a while loop with good manners and a strict tool schema. #AIAgents #CodingAgents
Framework lock-in is where enterprise AI agent projects go to die. When platform teams mandate an opinionated SDK, teams using LangChain or custom stacks get boxed out of deployment and eval pipelines. Today Google opened up the Agents CLI Extension System: bring any framework, keep unified Cloud Run deployment.
5 of the 14 posts behind this answer, one per person. The other 9, and every new one as it’s posted, come with tracking. Get early access
Questions· 17 posts from 14 people· written 27 Oct
What do people ask about LangChain?
Three questions are level at 3 people each: whether it is still the way to go, how the usage and benchmark numbers are measured, and how an agent's auth works. Close behind, 2 people ask whether the defaults cover the unsafe parts — timeouts, dry runs, and paused runs nobody answers.
The currency question is asked plainly: is LangChain and LangGraph "still the way to go" for agentic workflows or is there something with less hassle, are they actually outdated when enterprise teams still use them, and — from someone who once proposed it at work — what happened to them, "now I don't even hear about it".
The measurement questions press on denominators: does a call footprint count retries, since a harness re-calling after a failed structured-output parse books calls nobody chose; how are the two measures defined and are retries or multi-model routing counted separately; and was the eval set locked before tuning, since otherwise topping your own benchmark says little.
Auth questions ask for the boundary: scoped credentials per session, tenant-isolated memory, traceable tool access, whether those show in traces, and how the agent acts for a specific user versus a shared service account. On defaults, people ask whether tool-call timeouts are shipped, whether anyone uses a dry-run log first, and what happens when a paused run gets no answer for hours. Smaller asks: where Managed Deep Agents fit and what real tasks run on them, how to divide orchestration from evaluation when using DeepEval alongside, and whether it works from JavaScript.
- Is it still the way to go?3
- How are these numbers measured?3
- How does agent auth work?3
- Do the defaults include timeouts and dry runs?3
- What can Managed Deep Agents do?2
- How do people wire it up in practice?2
See 5 of 17 postsHide the posts
礼貌提问 Langchain、 LangGraph 真的过时了吗？ 现在企业应用开发还是用的比较多吧？
Langchain and Langgraph is still the way to go for building agentic workflows, right? Or are there any good alternatives with less hassle?
does the call footprint count retries? if it does, part of that ranking is reliability, not preference: a model that trips a structured-output parse gets re-called by the harness and books calls nobody chose.
auth is where most agent demos quietly die. curious how you handle the agent acting for a specific user vs a shared service account in the loan flow
half the agent bugs people post about here are just missing timeouts tool call hangs, agent waits forever, nobody notices till the bill or the angry email put a timeout on every tool call before you touch the prompt. @LangChain does yours default to one?
5 of the 17 posts behind this answer, one per person. The other 12, and every new one as it’s posted, come with tracking. Get early access
Wishes· 8 posts· written 27 Oct
What do people want from LangChain?
The biggest ask, from 2 people, is feedback in Slack: a reaction showing an agent accepted the request, because silence between request and answer reads as broken. Everything else is one person's ask, from signed authority receipts to traces carrying tenant and permission context.
The Slack ask is about the gap, not the answer: workflows "feel broken in Slack simply because there's no feedback between the request and the final response", and a reaction is called a lightweight fix. One post describes the shipped version — an agent that posts 👀, switches to 🐛 when a message says "broken", and falls back to 👀 when its classifier is under 25% sure — which is where the request and the release meet.
The single asks are specific about what they want recorded. One wants an authority receipt — agent id, issuer, scope, policy version, approval rule, expiry, denial reason, audit URL — so agents can refuse unsafe calls before authority leaks. Another wants traces to carry run IDs, tenant and permission context, and model and prompt versions, without retaining sensitive payloads. An eval should separate "didn't retrieve it" from "retrieved it but missed the answer", since the fixes differ. Others ask for a demo dataset that is a real 14-year-old inbox, an approval step before the run rather than a reject button on a finished diff, and support for other programming languages — from someone who would not use anything else.
- Reactions to acknowledge requests in Slack2
- Authority receipts for agent actions1
- More context in traces1
- Eval that separates retrieval from answering1
- A realistic demo dataset1
- Support for more languages1
See 5 of 8 postsHide the posts
I’ve seen agent workflows feel broken in Slack simply because there’s no feedback between the request and the final response. Reactions seem like a lightweight way to fix that.
【SlackのAIエージェントが絵文字で依頼の受付を知らせる仕組み】 LangChainは公式ブログで、AIエージェントの構築・運用環境「Managed Deep Agents」に追加したSlackのリアクション機能を紹介した。返信を待つ利用者に、依頼を受け付けたことを絵文字で知らせる。 絵文字は名前を表す文字列か、メッセージに応じて名前を返す関数で指定する。記事の例は通常👀を付け、「broken」が含まれる場合に🐛へ切り替える。Slackに登録した独自絵文字も名前で指定可能。 判定モデル（入力を分類して候補を選ぶモデル）を使う別の例では、各絵文字を使う状況を説明し、内容に合う候補を選ばせる。選んだ候補への確信度が25%未満なら👀に戻す分岐を関数内に設け、通常の絵文字で受付を知らせる。
LangChain needs an authority receipt: agent id, issuer, scope, policy version, approval rule, expiry, denial reason, and audit URL. Agents can refuse unsafe calls before authority leaks.
OpenTelemetry makes cross-provider traces much easier to compare. Including run IDs, tenant and permission context, plus model/prompt versions, would make failure reproduction stronger without retaining sensitive payloads. #AIAgents #Observability
Interesting!! I’d want an eval that separates ‘didn’t retrieve it’ from ‘retrieved it but missed the answer.’ They're very different fixes, and answer accuracy alone hides that.
5 of the 8 posts behind this answer, one per person. The other 3, and every new one as it’s posted, come with tracking. Get early access
In their words
How do people talk about LangChain?
The phrases people use most, counted once per person.
- 1“deep agents”5
- 2“context engineering”4
- 3“managed deep agents”2
- 4“tool calling”2
- 5“coding agent”2
- 6“platform teams”2
Posts are shown the way X shows them. One of them yours? Here’s why it’s shown, and how to take it off.
Also in LLM observability and evals
Get every post about LangChain, filtered and labelled
This page is the summary. Tracking gives you the stream behind it for any brand, keyword or @mention: each new pain point, request and question as it’s posted, with the spam and promotions already taken out.