Running on agentic reasoning and espresso. Building with @LangChain 🦜🔗.
San Francisco
- Tweets 3,107
- Following 899
- Followers 3,333
- Likes 3,543
Pinned Tweet
We’re launching Managed Deep Agents today 🎉👀
langchain.com/blog/managed-d… @LangChain
The goal posts for building agents have moved.
Giving an agent tools, deploying it somewhere, and putting a UI in front of it isn’t the hard part anymore.
The next challenge is everything around the agent: identity, memory, credentials, permissions, and securely connecting it to the services your users already use.
This beta release is just the beginning and our foundation what is about to come next 🚀
5
10
27
7,961
The simplest always-on agent is one that runs on a schedule ⏰
Two ways in Managed Deep Agents:
1️⃣ You decide: add a file to schedules/ and mda deploy turns it into a cron job
2️⃣ The agent decides: a tool calls runtime.schedules, so users can say "remind me Friday at 9"
3
4
263
Docs, including one-shot runs, Slack delivery and thread options:
docs.langchain.com/langsmith…
Python: docs.langchain.com/langsmith…
1
54
Christian Bromann retweeted
🔎 Announced at Interrupt: AI Search in LangSmith
You used to need an evaluator or read hours of conversations to find traces that answer questions such as:
💬 Does the user seemed frustrated?"
💬 Where did the agent fail to complete the task?
💬 Where is the user trying to cancel their subscription?
Now, just describe it. LangSmith returns the traces where it happened. langchain.com/blog/agent-tra…
20
12
75
27,202
Christian Bromann retweeted
Love the debugging UI in @Cypress_io or @playwrightweb? It’s now in WebdriverIO too 🤩
Every command, the page after each step, and your source, live. Mobile tests included. CI saves a trace you can replay anywhere.
Contributed by our partner @browserstack 🧡
webdriver.io/docs/devtools
2
1
8
617
Giving your agent a browser is the easy part. Getting it to use one reliably is the hard part.
With Managed Deep Agents:
👉 A sandbox with Chrome and wdio session
👉 Skills in plain Markdown with the recipe, so the agent doesn't guess
No custom tools. Code 👇
1
8
310
🧑💻 github.com/christian-bromann…
The @Google flights skill is the recipe: open a "?q=" search URL, switch to Cheapest, and read every fare from the page's aria-labels in one "wdio session exec" call. No screenshot-and-guess loop. One SF to Berlin comparison cost $0.07 in model calls, with the screenshot inline in the LangSmith trace.
Docs for wdio session: webdriver.io/docs/session
71
Published @webdriverio v10.0.2 with improvements to "wdio session" that introduces @wdio/snapshot which handles snapshotting of web pages and native apps across @wdio/cli and @wdio/mcp pushing "wdio session" to the top of our benchmark leaderboard 👑
benchmark.webdriver.io
1
13
906
Christian Bromann retweeted
WebdriverIO v10 is here 🎉
Built for agent verification loops: your coding agent opens your app in a real browser, on iOS or Android, or as a desktop app, checks the feature works, and leaves a test behind.
Works with @claudeai, @cursor_ai, @OpenAIDevs and @GitHubCopilot.
5
3
23
2,043
Christian Bromann retweeted
managed deep agents 0.9 is here! things we shipped:
⏰ agents can schedule tasks OBO themselves or the user
🏭 dynamically configure your mda per run (models, tools, skills, etc.)
👀 Slack reactions show when your agent has seen a message
📁 improved multimodal capabilities
🔌 MCP 2.0 support via langchain.mcp
released for js and py langch.in/mda
4
7
23
3,088
I benchmarked 7 browser setups for coding agents on real tasks on live websites.
With Claude Sonnet 5:
- @Stagehanddev: 61%
- wdio session: 58%
- @webdriverio MCP: 54%, cheapest per task and token
- agent-browser 50%
- @playwrightweb: 46%/37%
benchmark.webdriver.io
In v10 `wdio session` lets your agent drive any app from the shell:
open chrome localhost:3000
snapshot
click e3
export --out cart.e2e.ts
We benchmarked it against 6 other agent browser setups on real tasks on live websites. It tied for the top success rate with Claude Sonnet 5.
4
1
9
474
I tried to keep it fair:
- same harness, prompt and model for every setup
- tasks from Online-Mind2Web, picked by none of us
- a judge from another model family grades every run
Caveat: 100 tasks is still a small sample, and the leaders are effectively tied. More runs and models coming. This did cost me ~$600 already 🙈
1
3
84
Best part so far: going through every task wdio session failed showed gaps in how it presents long pages, frames and custom widgets. Fixing them took it from 22 to 58% and up.
All open source. Think we set your tool up wrong? Open an issue:
benchmark.webdriver.io
2
68
I am so excited about this release @webdriverio and every single new feature 😊
If a task is verifiable, agents can do it. A lot of people are noticing that right now.
So the question isn't "can the agent write the feature?" It's "how does it prove the feature works?"
🧵 1/3
WebdriverIO v10 is here 🎉
Built for agent verification loops: your coding agent opens your app in a real browser, on iOS or Android, or as a desktop app, checks the feature works, and leaves a test behind.
Works with @claudeai, @cursor_ai, @OpenAIDevs and @GitHubCopilot.
2
2
10
667
For a web or mobile app, that proof looks like this:
- open the real app, not a mock
- click through the feature like a user would
- check the screen and the network
- leave a test that proves it on every change
I really enjoyed reading @addyosmani's piece "The Code Nobody Reads" which calls the system around this loop engineering.
1
3
133
A loop is only as good as its check. That check is what we built @webdriverio v10 for: Chrome, Firefox, Safari, iOS, Android, Electron and desktop apps, with a plain test at the end.
191 contributors made it happen:
webdriver.io/blog/2026/10/05…
3
138
Christian Bromann retweeted
@𝚕𝚊𝚗𝚐𝚌𝚑𝚊𝚒𝚗/𝚖𝚌𝚙-𝚊𝚍𝚊𝚙𝚝𝚎𝚛𝚜 2.0 is out! 🧵
Using MCP with your TypeScript agents up to MCP servers just got a lot simpler: support for the latest stateless version of MCP, tools that can check in with your users when they need something, much easier auth integrations, and a ton of QoL updates.
Here's what's new 👇
7
8
28
14,017
This makes me really excited about the future of LangGraph and the @LangChain ecosystem at large 🤩 langchain.com/blog/building-… 🚀
4
209
Christian Bromann retweeted
managed deep agents feels like magic. it handles the entire slack setup for you, and if you haven't picked an icon, it derives one from the agent's name. ~2M possible icons! inspired by dither-kit's DitherAvatar, thanks @grimcodes. langch.in/mda
5
10
45
16,368
Christian Bromann retweeted
the claims runtime in openwiki v0.4.0 gives your wiki a persistent source of truth to help it forget and self-correct.
but how would a reader know:
1⃣ what sources a page was built from?
2⃣ if a page is fully verified?
3⃣ who generated the page in the first place?
to answer that we migrated openwiki to support OKF v0.2
the claims runtime powers page-level sources and verification and openwiki separately stamps who generated the page and when.
claims=content-level truth
okf=page-level trust
special thank you to @seo_dev2142 for kicking off okf v0.2 support in openwiki!
8
14
89
37,497
Starting today, @LangChain’s Managed Deep Agents can bake your agent’s environment at deploy time. 🧑💻
No clone. No install. No repeated setup.
Drop in setup.sh or a Dockerfile. Bake once. Every new thread starts ready. 👇
2
5
18
13,379
If sandbox/setup.sh exists, mda deploy and mda dev run it once and save the result as a snapshot.
New threads clone that snapshot instead of cloning the repo and installing packages again. The snapshot is reused until setup.sh changes.
Then every agent starts on a branch, not in setup.
1
2
246
You can also start the bake from a published Docker image, or from an existing LangSmith snapshot.
One sandbox per thread. Secrets stay in env, not on disk, because anything setup.sh writes is part of every later clone.
Docs: docs.langchain.com/langsmith…
1
151