AI Daily · 2026-10-10

Deno is joining Cloudflare: the Deno runtime will continue receiving monthly bug fixes and security updates for another year before official development ends, while the project remains open source; Ryan Dahl said the decision was mutual and argued Deno had been pulled into Node compatibility's gravity well. OpenAI's Asana case study shows GPT-6 Astra helped optimize a browser agent so that on GPT-6.1 Sol estimated model costs dropped 76x and speed rose 5x, largely by extending caching to browsing history and screenshots. LangChain published two pieces—one on designing loading states and reactions for Managed Deep Agents, another on building a model router inside an agent harness that cuts median thread cost by 64%. Standard Bots also shared its industrial robotics AI stack after a $200M Series C, using a shared foundation model tuned per customer, with cloud training and local inference.

North America · First-hand

OpenAI

⭐⭐ [Other] Asana cuts model costs 76x in browser tests with GPT-6.1 Sol

OpenAI published an Asana case study: using GPT-6 Astra in Codex to run experiments, Asana optimized its browser agent's workflow on GPT-6.1 Sol to run 76x cheaper and 5x faster in tests. Asana's 144-run study tested GPT-6.1 Sol and three other frontier models (referred to as Models A, B and C); the optimized workflow averaged $0.47 in estimated model costs and about four minutes per run, 76x cheaper and 5x faster than the original production setup on Model B. GPT-6 Astra found the agent cached its fixed instructions and tool definitions but not the growing history of page text and screenshots, so every request resent that history at full price. Hidalgo selected three fixes to test: extending caching to the browsing history, increasing the amount of retained text, and removing screenshots in batches rather than at every step; the best-performing policy let screenshots accumulate to 20 before cutting back to the most recent one. Work he estimated would take one to two months by hand took about a week with GPT-6 Astra.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Product Update

⭐⭐ [Product Update] How to build great out-of-the-box user experiences with Managed Deep Agents

LangChain published a tutorial on building out-of-the-box user experiences for agents with Managed Deep Agents, focused on designing loading states and receipts. In v0.9, Managed Deep Agents adds a reactions API for Slack channels that manages emoji responses; the reactions argument accepts either a string emoji or a callable that returns one. The example defaults to 👀 and returns 🐛 when a message contains the word 「broken」. Emoji selection can use simple heuristics or a decision model such as Jev, which picks from a predefined vocabulary based on input context (Jev supports up to 255 choices, and a decision matrix with multiple questions can extend the set). The post uses LangChain's internal marketing agent for video content as an example of why users need visibility into long-running agent tasks.

⭐⭐ [Product Update] How to Build a Model Router in the Harness

LangChain describes how it built a model router inside the agent harness, using its open-source coding agent Open SWE as the test bed. The post argues the routing decision belongs in the harness rather than a generic gateway, because picking a model requires the domain and task context the harness already assembles. Against a baseline that always used a top-tier frontier model, the router cut median cost per thread by 64% with no measurable change in quality. The build started by understanding the task space: the team labeled a week of interactive threads with an LLM classifier by task type, finding new features (22%) and bug fixes (17%) as the largest groups, followed by test or no-op runs (16%). The post also uses total cost and invocation count as signals of task complexity, and says effective router design rests on observability and evals.

Other

⭐⭐⭐ [Other] Deno is joining Cloudflare

The Deno team released the first version of celld in August, their open source implementation of the Durable Objects pattern from Cloudflare Workers; now Cloudflare is acquiring Deno outright, with the goal of building on celld to make workerd self-hosting a first-class supported way to build and run apps using the Workers programming model. The bad news is that Cloudflare will support the Deno runtime for another year with monthly bug-fix and security releases, after which it will end development of the Deno runtime; Deno will remain open source and others are welcome to continue it. Deno (and Node.js) creator Ryan Dahl explained the decision in a Hacker News comment: it is a joint decision he agrees with, and he no longer thinks Deno is where he can do the most important work — it has been pulled into the gravity well of Node compatibility, forcing it to behave exactly as Node does, while he is interested in building powerful new abstractions; celld, which depends only on object storage for coordination and persistence, is an entirely new model for server development. Simon Willison adds that his favorite Deno feature has long been its permissions system, which lets you specify exactly which files and folders a script can read and write and which network hosts it can access; Node.js added a similar permissions model in v20.0.0 in April 2023, declared stable in v22.13.0 in January 2025, but it does not yet support allow-listing specific network hosts.

⭐⭐ [Other] Building AI for Reliable Execution: Lessons From Industrial Robotics

Latent Space interviewed Evan Beard, co-founder and CEO, and Leif Jentoft, Head of AI, at industrial robotics company Standard Bots about its AI stack. The company describes itself as America's largest AI-native industrial robot manufacturer, counts NASA, Amazon and Lockheed Martin as customers, and recently raised a $200 million Series C at a $1 billion valuation led by General Catalyst and robotics-focused fund RoboStrategy. Its robot arms target tasks such as machine tending, welding and assembly: a shared base model is adapted by customers through demonstrations and fine-tuning, and the largest of its models is in the low billions of parameters. Jentoft says the company values data quality far more than raw volume, and that in-situ interventions from real deployments can fix an edge case with a few dozen examples. In a typical machine-tending routine the learned model handles perception — locating and identifying parts — while conventional programming handles motion and cell logic, with the zero-shot backbone trained on over a billion images. Training happens in the cloud while inference runs locally, and because the company controls the arm, end effector, control system and AI, it says it can co-optimize models and control policies.