WEBVTT NOTE The Rundown — nextbig.dev daily audio edition, 2026-03-24 1 00:00:00.000 --> 00:00:07.361 Good morning and welcome to the Builder's Briefing for March 24th, 2026. I'm Alex, here with Sam, and we've got a packed show — a viral open-source agent framework, some sobering real-world data from Walmart on AI checkout, and the Department of Defense apparently forgetting to renew a TLS cert. 2 00:00:07.361 --> 00:00:10.544 Love it when the cybersecurity folks can't manage their own certificates. But yeah, lots of good stuff today. Let's get into it. 3 00:00:10.544 --> 00:00:17.781 Alright, so our big story — Pascal Editor just dropped as an open-source platform for building autonomous agents, and it blew up overnight. We're talking over eleven thousand stars on GitHub basically right out of the gate. The repo is positioning itself as 'autonomous agents for everyone.' 4 00:00:17.781 --> 00:00:22.233 Eleven thousand stars overnight is wild. That's the kind of traction that usually takes weeks, even for well-funded projects. What's the actual pitch here — like, who is this for? 5 00:00:22.233 --> 00:00:29.096 So if you've been stitching together LangChain with custom tool-use logic and brittle orchestration code, Pascal Editor is basically gunning for that entire layer. It's a framework for spinning up agents that can take actions, use tools, and operate with minimal hand-holding. 6 00:00:29.096 --> 00:00:35.463 That's interesting because the 'for everyone' framing sets it apart from things like CrewAI or AutoGen, which honestly have a pretty steep learning curve. The agent framework space is getting so fragmented though — I feel like there's a new one every week. 7 00:00:35.463 --> 00:00:43.918 Exactly right. And the signal here isn't just about Pascal specifically — it's that developer experience is becoming the battleground. The winner won't be whoever has the most features; it'll be whoever nails composing, debugging, and deploying agents. Keep an eye on their plugin ecosystem over the next sixty days to see if this has legs. 8 00:00:43.918 --> 00:00:45.535 Good call. Stars are vanity — plugin adoption is the real metric. 9 00:00:45.535 --> 00:00:49.688 Okay, shifting to AI news, and this one's a jaw-dropper. There's a demo of the iPhone 17 Pro running a four hundred billion parameter LLM on-device. Locally. No cloud. 10 00:00:49.688 --> 00:00:55.432 Wait, four hundred billion parameters on a phone? That's — I mean, that completely changes the math on local versus cloud inference for mobile apps. A year ago we were excited about seven billion parameter models running on-device. 11 00:00:55.432 --> 00:01:00.779 Right, and what's wild is the implication — your users are going to expect offline-capable AI within a year. If you're building AI-powered mobile apps, you should be prototyping on-device workflows now, not waiting. 12 00:01:00.779 --> 00:01:04.460 The latency alone would be a killer feature. No round trip to the cloud, no API costs per request — that's a completely different product economics. 13 00:01:04.460 --> 00:01:08.613 Now here's the counterpoint story that I love. Walmart shared real data showing that ChatGPT-based checkout converted three times worse than their traditional website. 14 00:01:08.613 --> 00:01:10.627 Three X worse? Ouch. That's not a marginal difference, that's a catastrophic one. 15 00:01:10.627 --> 00:01:17.193 Total cold shower for the conversational commerce crowd. Users still want structured UI when they're actually trying to buy something. Chat is great for discovery and browsing, but when it's time to convert — people want buttons, forms, the familiar checkout flow. 16 00:01:17.193 --> 00:01:22.987 That makes intuitive sense when you think about it. High-intent actions need precision and speed. Nobody wants to type 'yes, confirm my order, yes that's the right address, yes use my Visa' in a chat window. It's more work, not less. 17 00:01:22.987 --> 00:01:29.254 Exactly. And there's also a great post trending about someone who built an AI receptionist for her brother's mechanic shop. Documented the whole process. The takeaway was that the hard parts weren't the AI — they were the business logic and edge cases. 18 00:01:29.254 --> 00:01:34.427 That's always the story, isn't it? The demo is easy, production is where reality hits. Link in the briefing for that one — it's a great case study if you're building voice or chat agents for small businesses. 19 00:01:34.427 --> 00:01:40.967 On the developer tools front, Antithesis — the deterministic testing company — released Bombadil, which brings property-based testing to browser UIs. If your frontend test suite is all happy-path snapshots, this could catch the weird state bugs that slip through. 20 00:01:40.967 --> 00:01:45.891 Oh, property-based testing for the frontend is something I've wanted for years. Complex forms, multi-step flows — that's where the gnarliest bugs live and where snapshot testing just doesn't cut it. 21 00:01:45.891 --> 00:01:51.511 Also worth flagging — tinygrad keeps gaining traction on GitHub. George Hotz's minimal ML framework. If you need something that isn't PyTorch, especially for edge deployment or custom hardware targets, that's the one to watch. 22 00:01:51.511 --> 00:01:54.720 Tinygrad is one of those projects where you read the code and actually understand what's happening. That's rare in ML frameworks. 23 00:01:54.720 --> 00:01:59.146 Okay, this one made me laugh. Cyber dot mil — the Department of Defense's cybersecurity resource site — has been serving downloads over an expired TLS certificate for three days. 24 00:01:59.146 --> 00:02:06.060 The irony is just perfect. The literal cybersecurity arm of the DoD can't keep their cert current. But honestly, this is a reminder for all of us — automate your cert renewals. If you're still doing manual TLS, you are one missed calendar reminder away from being the next meme. 25 00:02:06.060 --> 00:02:11.705 Also on the infrastructure side, GitHub's uptime has been slipping below three nines, with multiple recent outages. If your CI/CD is fully GitHub-dependent, build some fallbacks — cache your dependencies, mirror critical repos. 26 00:02:11.705 --> 00:02:16.505 That's a big deal. A lot of teams just assume GitHub is always up, but if your deploys can't happen because GitHub is down, that's a single point of failure you should have addressed yesterday. 27 00:02:16.505 --> 00:02:21.553 Quick hits — POSSE, which stands for Publish on Your Own Site Syndicate Elsewhere, is resurfacing on Hacker News with over three hundred and sixty points. The own-your-content movement is alive and well. 28 00:02:21.553 --> 00:02:25.631 Love that. Also saw that PC Gamer wrote a thirty-seven megabyte article recommending RSS readers. The web bloat discourse continues to be unintentionally hilarious. 29 00:02:25.631 --> 00:02:32.769 There's also a RollerCoaster Tycoon optimization deep-dive — assembly-level performance lessons that honestly still apply to modern work. And a practical guide to migrating to the EU is trending hard with over six hundred sixty Hacker News points. Links for all of these in the briefing. 30 00:02:32.769 --> 00:02:40.006 So here's the big takeaway from today. Two signals that cut against each other, and both matter. Pascal Editor's viral launch shows the demand for simpler agent frameworks is enormous. But Walmart's data proves that agents replacing structured UIs in high-stakes flows is still a losing bet. 31 00:02:40.006 --> 00:02:45.278 Right, so the move is — use agents for discovery, triage, and orchestration. Let them do the fuzzy, exploratory stuff. But when it's time to convert, when the user has high intent, hand off to a purpose-built UI. 32 00:02:45.278 --> 00:02:50.500 The builders who nail that handoff — the moment where the agent passes the baton to a structured flow — they're going to outperform both the 'agents for everything' crowd and the 'agents are all hype' skeptics. 33 00:02:50.500 --> 00:02:53.534 It's not about agents versus UI. It's about knowing when each one is the right tool. That's the nuance everyone's missing. 34 00:02:53.534 --> 00:02:57.637 Well said. That's your Builder's Briefing for March 24th. Check the show notes for all the links — there's a ton of good stuff we didn't have time to cover in depth. 35 00:02:57.637 --> 00:03:00.000 Go clone that Pascal repo, automate your TLS certs, and we'll see you tomorrow. Happy building.