Agents don't just need access to services. They need access to where users actually work. Browsers. Apps. Devices. Local state. A look at how headless tools bring those capabilities directly into the agent loop.

Don't loose your customer when they switch between application views. useStream reconnects to your thread mid run and reconciles state and messages outside of runs, making UI thread management easy to manage 😊 Docs: docs.langchain.com/oss/javas…

Streaming in LangChain/LangGraph now has one mental model across the stack: start a model, agent, or SDK stream, then read the same projections: .text, .reasoning, .output. Same async iteration, same final value. One flow, everywhere. That's the win 🥇

A sweet example of what you can build with @LangChain streaming: a tiny multimodal bedtime story app where the story, illustrations, and narration audio arrive as they are generated. Built with @langchain/react media hooks LangGraph SSE streaming 🚀

ICYMI: You can now give your agents an in-memory TypeScript interpreter to enable some novel agent capabilities. Built using oxc and quickjs. See this writeup on how to use it!

your daily reminder that open models are plenty capable for a lot of coding work. easiest place to feel that out is deepagents! swap the model and go. i've been enjoying GLM-5.1, Kimi K2.6, MiniMax M2.7, DeepSeek V4 Pro. here's some examples using our CLI agent in headless mode

small workflow note that adds up. /staged-pr is a skill (via slash command) i run when wrapping up a PR. it takes my staged code changes and drafts a concise PR title & description, based on my preferences and our repo conventions. it's routine, formulaic, and well-scoped. i hit it 20 times a day. in the Deep Agents CLI i use /model to swap mid-session: heavier frontier model for the actual coding, then over to glm 5.1 (via @OpenRouter ) or kimi 2.6 (via @baseten) for the skill. it yields indistinguishable quality differences, runs faster, and is ~5x cheaper than leading LLMs. the broader point: matching the model to the task beats picking one model for everything. open models are extremely good at the long tail of routine agent work, even if frontier still wins the hard stuff. a lot of what we point frontier models at isn't actually that hard! -- linking skill below for those interested in trying it out

we're continuing to see clear examples where a model's harness is a major determinant of overall performance. with the same model, running on same task, it's easy to observe very different scores depending on (system) prompts, tools (& their descriptions), and middleware (steering hooks). this is exactly why we built a harness profiles abstraction in Deep Agents: per-provider or per-model overrides for base system prompts, tool names implementations, etc., so swapping models doesn't mean losing the work that made the last one good! 10–20pt jumps on tau2-bench in our own testing. currently cooking up built-in profiles for popular open weight models 🧑‍🍳 langchain.com/blog/tuning-de…

the deeper point: model choice and harness choice aren't independent variables. benchmarking a model without specifying the harness is like benchmarking a chip without specifying the compiler. the number means something, but not what people think it means. profiles make the harness a first-class, named object — you can diff it, version it, and swap it when you swap models.