内容历史
come join our team and make brace happy!
if this isn’t how you react to new hires joining your team, I fear you’re not hiring correctly
RT Brace if this isn’t how you react to new hires joining your team, I fear you’re not hiring correctly Original tweet: https://x.com/BraceSproul/status/2011640219292831979
great use case for agent builder! an ambient agent to monitor slack channel and create linear tickets
Docs requests kept getting buried in Slack 🫠 so I built an agent with LangSmith Agent Builder that creates Linear tickets, pulls in context, links back to threads, applies labels, and auto-assigns. No code, just a prompt! ✨ Try it for yourself: https://smith.langchain.com/agents?skipOnboarding=true
RT Lauren @ LangChain Docs Docs requests kept getting buried in Slack 🫠 so I built an agent with LangSmith Agent Builder that creates Linear tickets, pulls in context, links back to threads, applies labels, and auto-assigns. No code, just a prompt! ✨ Try it for yourself: https://smith.langchain.com/agents?skipOnboarding=true Original tweet: https://x.com/docs_plz/status/2011536177556570203
We get asked a bunch about multi-agent architectures, usually when teams want to scale and combine multiple specializations in one cohesive experience @sydneyrunkle wrote some guidance on when to use which patterns (not all are multi-agent!) https://www.blog.langchain.com/choosing-the-right-multi-agent-architecture/
RT Tadeo Donegana Braunschweig I’ve been using @LangChain Agent Builder since the beta, and now that it’s officially launched I wanted to share how I’ve been using it in practice and some feedback. cc: @BraceSproul @hwchase17 Original tweet: https://x.com/tadeodonegana/status/2011452497647595733
RT Kevin Frank Re @LangChain cares deeply about shipping reliable agents, and Agent Builder does just that. In just a few minutes, you can create an agent tailored to your use case. And you can actually understand what's going on under the hood. In Agent Builder, you can see: • Agent: the system prompt that defines the agent's role and instructions • Toolbox: the tools the agent can actually call • Sub-agents: how complex tasks are decomposed and delegated • Skills: reusable capabilities saved to memory (so you don't re-prompt every time) And it integrates directly with LangSmith for observability and evaluation of your agent. To increase agent adoption, we need people to understand everything going on under the hood. Agent Builder and LangSmith do that. Original tweet: https://x.com/KevinBFrank/status/2011154462128144539
✒️How I built an AI agent to automate my emails with LangSmith Agent Builder LangSmith Agent Builder is a no-code agent builder. I built an email assistant to monitor and respond to emails, that I've been using for the last ~3 months. Here's what it looks like: 1/ Triggers: it
RT Brace OG Agent Builder tester ❤️👇 Original tweet: https://x.com/BraceSproul/status/2011145199951757762
I've been lucky enough to have access to @LangChain Agent Builder for a few weeks now and have gotten to know it pretty well This isn't some drag-n-drop agent builder Deep Agents that are anything but basic - Customized with your own tools - BYOMCP 🚀 Comparison Matrix Agent
RT Git Maxd Deep Agents Builder in LangSmith is so easy to use and so powerful you’ll have to do a double-take 👀 Tell the @LangChain Agent Builder a rough idea what you want, iterate in natural language, tune by hand in the Editor, immediately connect via API and Trace 🚀 Epic Intro! 👇 Original tweet: https://x.com/GitMaxd/status/2011131633051763138
✒️How I built an AI agent to automate my emails with LangSmith Agent Builder LangSmith Agent Builder is a no-code agent builder. I built an email assistant to monitor and respond to emails, that I've been using for the last ~3 months. Here's what it looks like: 1/ Triggers: it
RT Brace Agent Builder is incredibly valuable for anyone, not just people who don't know how to code. See below where Harrison shows how he automated his email using Agent Builder (i can confirm Harrison knows how to code, but still chose to use Agent Builder for this!!) Original tweet: https://x.com/BraceSproul/status/2011126819681091640
✒️How I built an AI agent to automate my emails with LangSmith Agent Builder LangSmith Agent Builder is a no-code agent builder. I built an email assistant to monitor and respond to emails, that I've been using for the last ~3 months. Here's what it looks like: 1/ Triggers: it
✒️How I built an AI agent to automate my emails with LangSmith Agent Builder LangSmith Agent Builder is a no-code agent builder. I built an email assistant to monitor and respond to emails, that I've been using for the last ~3 months. Here's what it looks like: 1/ Triggers: it is triggered by incoming emails. I don't have to do any work to kick it off - it just runs automatically 2/ Tools via MCP: connects to gmail (read emails, send email) and gcal (read calendar, read events, create event) 3/ Human in the loop: the "write" actions (sending email, creating calendar) require human approval to run. More on this later - but wanted to highlight that it's able to go completely wild! 4/ Subagent for calendar scheduling: LLMs suck at working with calendars! So i have a subagent specifically for finding my availability - its works a lot better 5/ Agent inbox to review: as mentioned, some actions require human approval. LangSmith Agent Builder ships with an agent inbox to review and approve the actions the agent wants to take 6/ message_user to ask questions: sometimes my agent doesn't know what it should do. It has a message_user tool, which it can use to ask me a question! This also shows up in agent inbox 7/ Remembers what I say: it updates it memory automatically based on my responses to it! This keeps me from having to repeat myself Overall - I never look at my actual email anymore, just this! I recorded a quick video about how I made it + how I use it: https://youtu.be/bzcAZJTxOrs We made this into a template - so you can try it out easily here: https://smith.langchain.com/agents/templates?viewingTemplateId=email-assistant&skipOnboarding=true And if you want to build your own agent - try out LangSmith Agent Builder here: https://smith.langchain.com/agents?skipOnboarding=true
RT Christian Bromann Very neat hardware use case using @LangChain and voice 🗣️😊 Original tweet: https://x.com/bromann/status/2010846727071465764
M5Stack CoreS3 SEをPythonで自在に実装できるAIエージェントのフロントにするべく、PC上のPythonサーバにWebSocketでアクセスして、音声の入力と認識、音声合成して発話までつくることができた。一旦おうむ返しにしているが、コアの部分はPC上のPythonなのでLangChainとか使って作り込める。
RT Viv deepagents is batteries included so you get something good working asap but it’s fully customizable so you can cook from there like Jack 🫡 Original tweet: https://x.com/Vtrivedy10/status/2010397690975429083
I was worried that using a batteries-included harness like @LangChain's deepagents would take the fun out of building agents. It certainly didn't! I spent my Saturday night, comparing a freethinking deep agent with a heavily orchestrated agent workflow. Both tasked with writing
RT Jack Mu I was worried that using a batteries-included harness like @LangChain's deepagents would take the fun out of building agents. It certainly didn't! I spent my Saturday night, comparing a freethinking deep agent with a heavily orchestrated agent workflow. Both tasked with writing me a curated newsletter. Here are my big takeaways comparing a long-running agent with an orchestrated agent: 1. It's both shocking and fun to see how LLMs solve complex tasks 2. Orchestration is doing the reasoning for an agent. An agent in a loop can rack up tokens (just experimenting can cost a few dollars in API cost from Anthropic). An orchestrated workflow allows you to surgically decide LLM inputs and outputs -> agents are decisive, without as many turns and LLM calls. 3. Long-running agents create their own orchestration plan. They're more flexible than orchestrated workflows, and part of the harness's job is to make sure their generated orchestration is ~similar to how a developer would orchestrate a workflow. So which agent writes a better custom newsletter? The freelancing agent or the structured agent? I liked the free agent, but you can read them in my blog to decide for yourself :) (@Vtrivedy10 had to finally try deepagents for myself!! Doing this made me think of your agents as a workflow builder post) Original tweet: https://x.com/jackmuva/status/2010391908372324766
RT Seth Wieder Re Created a `pm-agent` repo Analysis - Synthesize feature requests + identify gaps in roadmap - Cluster user's NL search logs (incl. LangSmith traces) Reflection - Weekly x-team execution tracking - Monthly OKR review - Quarterly career growth reflections How it might manifest outside CLI: Original tweet: https://x.com/next_gen_seth/status/2010207690354520367
Love this. This is the future of software. Delegating a team of agents, doing work on your behalf, with scheduled workflows, and a centralized inbox for command and control.
View quoted postLet a millions harnesses bloom
launching nanocode! minimal claude code implementation. zero deps, ~250 lines of python. full agentic loop with tools (read, write, edit, glob, grep, bash). prompt is just "concise coding assistant. cwd: /path"
RT Viv subagent = agent + harness design on how to execute as people build more agents, getting questions on "how do I use them together"? "I want to reuse my agent to do X specialized task!" this whole set of questions centers around "multi-agent orchestration"....and there's lots of design decisions. but start basic (including don't use them!) Subagents are the first google result on this so start there: 1. subagents are just agents (they’re literally defined by the same files) --> agent files 👀 2. the “sub” is in the harness decision on how the agent is executed. Sync vs async, resumable, how to pre-load context, start prompt, how to call them 3. most common pattern is via a Task tool. Subagents get one turn, send a final message back to the orchestrator agent and then self destruct. You could also offload their run to file system....design decisions it's not that scary...they're fun to play around with, and they're a great first step into multi-agent orchestration. i use them all the time with our deepagents library Original tweet: https://x.com/Vtrivedy10/status/2010091537737105773
RT Git Maxd Starting to feel that @LangChain Agent Inbox vibe Deep Agents runs surface the agents response to the inbox where you can reply to the Agent conveniently and check-in on things If you haven’t tried the LangSmith Deep Agent Builder yet, you should! Original tweet: https://x.com/GitMaxd/status/2010047378301346186
RT Afiz ⚡️ LangGraph 101: (with code snippets) Are you new to LangGraph? Curious about how to build AI Agents using LangGraph @LangChain? This thread is for you. A Thread 🧵👇 Original tweet: https://x.com/itsafiz/status/2009980787760763119
RT cole murray Original tweet: https://x.com/_colemurray/status/2009807520886296661
agent files agents are just defined by markdown/json files now system prompt: http://agents.md subagents: subagents/ tools: http://skills.md + mcp.json
View quoted postRT Millin Gabani Always fun to see the best teams working on same concepts as you. Agents are just files and folders. The world is truly about to change. Original tweet: https://x.com/trillhause_/status/2009793991185695202
agent files agents are just defined by markdown/json files now system prompt: http://agents.md subagents: subagents/ tools: http://skills.md + mcp.json
View quoted posttraces matter!
standard awesome Anthropic blog, I cannot stress the importance of this step enough —> pls look at the agent traces!!! TONS of value in understanding if an agent “failed a task” then: 1. how did it fail? (formatting error, logical error, went down wrong track in planning,
RT Viv standard awesome Anthropic blog, I cannot stress the importance of this step enough —> pls look at the agent traces!!! TONS of value in understanding if an agent “failed a task” then: 1. how did it fail? (formatting error, logical error, went down wrong track in planning, environment bug…) 2. did it actually fail? should you update your grading if it was wrong? 3. what were successful patterns vs failing patterns (stratify across pass and fail and see the steps) 4. link it back to your instructions and tools/skills. did you not give the right tool to do the task? Are instructions ambiguous? Agent building is a deeply iterative process and this is literally the most golden data you can get to hill climb on, the list of things to look at is massive at LangChain when we want to make Deep Agents better, we mine the traces with an agent guiding us through what happened. Traces are huge, let agents help you!! there’s tons of human intuitions that don’t get baked into the agent in your first pass of designing them, you’ll learn stuff about the agent as you observe it and it will get better as you tune the knobs (prompts, tools/skills, harness design like context management etc) there’s a beautiful loop between designing agents and making them better systematically, agent design and eval design are deeply coupled. it’s great to see more people getting into it if you’re into building agents or making them observable so you can make them better, always down to talk 👀 Original tweet: https://x.com/Vtrivedy10/status/2009724848482762966
New on the Anthropic Engineering Blog: Demystifying evals for AI agents. The capabilities that make agents useful also make them more difficult to evaluate. Here are evaluation strategies that have worked across real-world deployments. https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents
View quoted postRT Afiz ⚡️ Happy Friday, it’s learning day! 🎓 I just enrolled in the LangGraph Foundation course and plan to finish it by the end of the week. @LangChain The best part? You can join too, absolutely FREE! 🚀 Let’s learn together! Original tweet: https://x.com/itsafiz/status/2009519114461798682
RT Git Maxd Great turn out tonight at the Phoenix @cursor_ai Meetup. These local meetups are a lot of fun - meeting people of all ages and everyone is so eager to learn from each other. Spoke to a lot of people about @LangChain Deep Agents. Lots of ASU students working with LangGraph! 🦜 Original tweet: https://x.com/GitMaxd/status/2009479064550297729
RT Viv yesterday we launched Ralph Mode for our deepagents library which is our fully open harness framework (checkout the linked vid/repo!) today we’re releasing native support for Skills & Memory in the deepagents SDK. Here’s a sneak peak into what we have coming with harnesses, Ralph, and autonomously improving software development Ralph Mode is really a harness level design decision (continual looping and keeping memory with filesystem+git). These harnesses are incredibly valuable in designing great agents, we want to make it as easy as possible to build, hack on, and fully customize them….harnesses should be fully open! I’m really excited about native Skills and Memory because they allow Ralph to update his knowledge over time by “Skillifying” the progress he makes. You can think of a setting where many Ralph’s go and create hundreds of skills that serve as a knowledge bank for other Ralphs to use, all easily tracked in git. Or Ralph learns and corrects himself over time in his memory files. We have a lot of cool stuff coming in the next week on this, super excited to build with the community :) (shoutout to our friend @GeoffreyHuntley for Ralph) Original tweet: https://x.com/Vtrivedy10/status/2009295526974595519
🫡 Ralph Mode for Deep Agents 🫡 What if you could give an AI agent a task and just let it run forever? We built Ralph Mode to test this, built on Deep Agents. Ralph Mode loops an agent with fresh context each iteration, using the filesystem as memory. Just start it, walk away,
RT Palash Shah some stuff we’ve been working on that’ll change the way you interact with agents. Original tweet: https://x.com/palashshah/status/2009126172362461289