Limrun (@limrun) provides the infrastructure to leverage your favorite cloud agents for mobile development too. Today, we provide remote XCode, IOS & Android Simulators to Replit, Rork, Momentic AI, and many more. Congrats on the launch, @muvaff! ycombinator.com/launches/QSu…

We're making Shopify's mobile apps self-healing. Agents watch production 24x7 for new crashes, reproduce them, write and verify a fix, then generate a short video explaining the root cause and the fix. Much faster to digest than the usual wall of text. An engineer then just reviews and ships it.

I predict that Anthropic is going to announce that they rewrote Linux around March 2027. I.e. they make a completely new OS that they start using for a majority of their agent work.

Up until recently it was just me hacking on prototypes, but my company is starting to feel real. We are 4 people, have a real physical office, and we just closed a seed round with @AmplifyPartners – looking forward to share a bit more about what we are working on soon!

I don't want 1000s sandboxes, I want an infinitely dynamic computer

Today, we're excited to announce Replicas has raised a $4M seed round We've built the harness-agnostic cloud agent Backed by @ycombinator, @flexcapital, @founderscoop, @paulg, @zenorocha and more. To celebrate, everyone who signs up today gets $300 in free credits. Here's how it will feel:

Introducing Selfdrive. The AI growth team for your iOS app that gets you users while you sleep. It handles Apple Ads, ASO, SEO and UGC creators for you. All of it. superapp.dev/selfdrive

we also shipped e2e v0.20.0 ⚡ → e2e run --watch → @limrun integration → Improvements to decision executor

You’d think I’d be the last person to say “sandboxes are dead.” But I’ll be the first to tell you that sandboxes aren’t the whole story for agentic infrastructure. A lot of the loudest takes here come from people who haven’t built agents that operate at scale. The reality is less convenient: there’s no single tool that fits every task. What matters is knowing what to use, when to use it, and which tradeoffs you’re making. Everyone feels like they need a strong take. It has to be A or B. In the real world at real scale, it’s usually some combination of both. The real world is messy. Infrastructure at scale is full of constraints you don’t see in a demo. If you haven’t dealt with them, it’s easy to mistake what worked for your agent for what should work for everyone. At @e2b, we believe agents will increasingly make these infrastructure decisions themselves. They’ll choose the resources and environments they need for the task, with the underlying infrastructure handling orchestration, isolation, and control. The only way to survive in AI is to keep building for where things are going. By the time you’ve built for what agents can do today, they’re already capable of more. And the easiest and fastest way to be wrong is to think your solution is the one. Sandboxes are part of that future. There’s a lot more to build around and alongside them. That’s what we’re building at @e2b

We’re releasing Tomo Interest Groups. The video below is an actual Tomo Interest Group interaction that a user shared with us (names anonymized) As Tomo learns more about you, it places you in group chats with other Tomo users pursuing similar goals. When Tomo offered this to 100 users studying for the LSAT, 60% of them opted-in and got placed into group chats with other Tomo users. They then traded notes and coordinated group study sessions, facilitated by Tomo. As more users join Tomo than ever before, they contribute to a pool of people with shared intentions waiting to discover each other. We’re rolling this out to more Tomo users with proven interests in hobbies, sports, careers, and side quests, because people go further together.

We are releasing React Native Skia version 3.0 which runs WebGPU by default. By moving away from the OpenGL-centric design, it allows for multithreading by default as well as deep integration with three.js, vision camera, and transformer.js youtube.com/watch?v=L-PNQi1n…

Introducing Replicas V3: the most flexible and cost-efficient cloud agent on the market. - Run Claude Code, Codex inside VMs - Bring any subscription or API keys - Delegate from Slack, Linear, desktop app, or your phone Now proactive, and more powerful than ever. replicas.dev

A humble attempt to est. the infra required to serve 100M DAU @Muse Rough conclusion is: 1 GW of power to serve 100M DAU in the base case, of which only ~0.1 GW comes from the CPU/VM layer. Depending on the # of reasoning-equivalent model calls one Muse DAU generates per day, 3-4GW is entirely plausible. Maybe that’s why @Meta is rumored to be adding 7-10GW of compute next year. The sandbox layer = sub $1B of CPU content and ~$2B of DRAM content, which is much smaller than many expected. Lot of moving assumptions. Welcome all feedbacks/ pushbacks. ------ Two very different pieces of infrastructure behind Muse. 1. Muse VM / sandbox infrastructure 2 vCPUs, ~8 GB of RAM and ~100 GB of persistent logical storage per user. starkinsider.com/2026/09/met… 2. Muse Spark inference Model inference goes out through Meta's external inference infrastructure. research.meta.ai/blog/securi… ------ 1/ Sandbox infrastructure A. CPU The first mistake is assuming that 100M DAU means 100M VMs are actively consuming compute at the same time. Suppose the average Muse DAU has an agent actively working for two hours per day. 100M users * 2 hours / 24 hours = ~8M average simultaneous active VMs Meta obviously cannot provision only for the daily average. Usage will be concentrated during waking hours and bursty. Assume a 2.5x peak-to-average ratio: 8M * 2.5 = ~20M peak active VMs Then add roughly 20% capacity headroom: ~25M provisioned live VMs. So the base assumption is effectively that Meta needs enough infrastructure to support roughly 25% of DAU being live simultaneously. The next important distinction is between virtual CPU allocation and physical CPU demand. Agent sandboxes are particularly well suited to CPU oversubscription. They spend a lot of time waiting. During those periods, the VM may still be alive, but it is barely using CPU. DeepSeek’s recently published DSec infrastructure provides a useful benchmark. Its production agent sandbox platform runs approximately 30,000 physical CPU cores and 250TB of DRAM across ~160 nodes, with peak concurrency above 380,000 sandboxes. arxiv.org/abs/2609.22978 DSec also demonstrates stable operation at around: 800 microVMs per node. With roughly 188 physical cores per node: 188 physical cores / 800 microVMs = ~0.23 physical cores per live VM. DeepSeek is obviously the King of efficiency. The number for Muse might be at 0.3-0.75 physical cores per live VM, or assume 0.5 physical cores per live VM as the base case. That is equivalent to roughly two simultaneously live Muse VMs per physical CPU core. Using the base assumptions: 25M live VMs * 0.5 physical cores per VM = 12.5M physical CPU cores. On a 256-core CPU: 12.5M cores / 256 cores per CPU = ~50K CPUs; Or on a 192-core CPU that would be 65K CPUs. At the current public pricing, that is ~$800M. B. DRAM CPU can be aggressively oversubscribed because a VM that is waiting may consume almost no CPU. Memory is harder to oversubscribe because a live VM still needs to retain its working state. Muse exposes roughly 8GB of RAM to the user environment, but one observed instance was actually using only around 3GB at the time of measurement. 25M live VMs * 3GB = 75PB of physical DRAM, call it ~75-100PB of physical DRAM feels like a reasonable base range. At the current public pricing, that is ~$2B. C. Sandbox power ~0.1 GW for the entire Muse sandbox / VM layer at 100M DAU. ------ 2/ Inference Muse’s personal computer executes tools and stores state locally, but the actual model runs on separate inference infrastructure. Meta’s Muse architecture Energy per inference event Microsoft’s 2026 study estimates that optimized frontier-scale inference consumes a median of approximately: 0.31Wh per normal query But a long reasoning query with roughly 15x the token count consumes approximately 13x as much energy, or around: 4Wh per long reasoning query The study specifically highlights reasoning and agentic workloads as significantly more energy intensive. microsoft.com/en-us/research… Sensitivity analysis on # reasoning-equivalent events per DAU per day Suppose each active @Muse user generates the equivalent of 50 heavy inference events per day. At 5Wh each: 100M users * 50 events/day * 5Wh = 25GWh/day 25GWh/day / 24 hours = ~1.0GW average power So inference alone could require: ~1-2GW of average power A 3-4GW Muse is entirely plausible. Maybe that’s why @Meta is rumored to be adding 7-10GW of compute next year. ------ The popular framing around Muse is that giving every user 2 vCPUs and 8GB of RAM creates an enormous CPU requirement. But the naive calculation materially exaggerates the CPU requirement because it treats logical VM allocation as dedicated physical infrastructure. The more interesting conclusion is: Consumer agents may be a meaningful new demand driver for CPUs and conventional DRAM, but inference remains the real compute bottleneck. And as agents do more work, run longer trajectories and increasingly spawn other agents, inference demand can scale much faster than the number of users itself. +++ Calling my peer review group: @bubbleboi @damnang2 @Midnight_Captl @FundaAI @fi56622380 . Feedback/ Pushbacks pls :). ++ Better formatted: robonomics.substack.com/p/ag…

One of the reasons we will win is because we are having more fun

Shape matters, too. iPhone Duo support and the new 3D view are available to use in our console and you can embed it in your web application with a single React component! Try it: console.limrun.com/stream?os…

🥳 This week we announced Agent Substrate. It's the compute runtime (that can work on Kubernetes) that enables everything you've seen from me. We can suspend and resume so fast that you can do it per tool call. 👉 github.com/agent-substrate/s… 👉 github.com/agent-substrate/e…

Workin on duo view 👀

cooking iPhone Duo👨‍🍳 @limrun

Greg Brockman says OpenAI pointed Astra at its own systems until it ran out of vulnerabilities to find: "We took 25% of our production engineers and said, 'Sorry, all your projects are on hold. You are now defending. You are now up-leveling our security architecture. You're going to use the models to find all the holes.' And we found a number of serious issues, and we fixed them." "We found some new problems, but eventually it saturated. We basically have found, to our knowledge, all of the P0s, all of the critical problems that Astra is smart enough to find. And of course, there will be a new model, there will be a new round." "You want to be in this tight loop of new cyber capability drops, you deploy it against your systems, you find the new holes, and ideally, you've managed to automate this, what we call defense factory. That's what we're building internally." "There are ideas, for example, formally verifying all of software, that are possible with AI." @gdb @bhorowitz

you can spin up a simulator for android and IOS apps directly in @tryreplicas by asking your agent to do so. this is a game changer when I want to test out new features in our mobile app. powered by @limrun