WEBVTT NOTE The Rundown — nextbig.dev daily audio edition, 2026-02-22 1 00:00:00.000 --> 00:00:05.678 Hey everyone, welcome to Builder's Briefing for February 22nd, 2026. I'm Alex, joined as always by Sam. Big show today — the entire ecosystem seems to have collectively decided that agents are real and they're shipping now. 2 00:00:05.678 --> 00:00:09.345 Yeah, it's one of those weeks where you look at GitHub trending and it's like, okay, the memo went out. Everyone got it. We're doing agents now. 3 00:00:09.345 --> 00:00:19.555 So let's start with the big story. Andrej Karpathy introduced this concept he's calling "Claws" — Simon Willison unpacked it on his blog, link in the briefing — and the idea is simple but powerful. It's a mental model for the tier of capability where AI agents stop just suggesting things and start actually reaching out and touching the world. Browsing, executing code, modifying files, calling APIs. 4 00:00:19.555 --> 00:00:27.423 Right, and what I love about this is it gives us shared vocabulary. Like, we've all been building these things, but we've been talking past each other. Karpathy's basically saying the LLM itself isn't the differentiator anymore — it's the claws. The tool-use interfaces, the sandboxing, the permission models. 5 00:00:27.423 --> 00:00:35.036 And the timing is perfect because look at what shipped this week. Google's Gemini CLI, Block's Goose agent, Cloudflare dropped an agent deployment framework, Expo shipped agent skills for React Native, Microsoft has an agent framework, and there's Cord for multi-agent coordination. All in one week. 6 00:00:35.036 --> 00:00:42.089 That's wild. And they're all slightly different takes on the same problem. Goose from Block is interesting because it goes way beyond autocomplete — it installs things, executes, edits, tests — and it works with any LLM. It's a true general-purpose coding agent you can extend. 7 00:00:42.089 --> 00:00:49.142 And then Cord tackles something most people are just starting to hit — how do you coordinate trees of agents working on decomposed tasks? That's the hard orchestration problem. If you've been building multi-agent systems and hitting coordination walls, that one's worth a look. 8 00:00:49.142 --> 00:00:56.628 So Karpathy's framework really nails it. The competitive surface for the next six months isn't prompt engineering — it's agent permission models, tool registries, and orchestration layers. MCP is becoming the de facto standard, and Google just shipped an MCP Toolbox specifically for databases. 9 00:00:56.628 --> 00:01:03.222 Exactly. If you're still building pure chat interfaces, you're building last year's product. Full stop. Alright, let's pivot to some broader AI news. Google's VP Darren Mowry basically issued a formal death notice for thin-wrapper startups and AI aggregators. 10 00:01:03.222 --> 00:01:10.173 Yeah, his message was pretty blunt — if your moat is "we call the API and add a UI," that's not a moat, that's a feature the foundation model provider is going to ship next quarter. You need proprietary data, deep workflow integration, or serious vertical domain expertise. 11 00:01:10.173 --> 00:01:17.074 And here's a related trend that's a little unsettling — there's analysis showing that basically every AI assistant company is converging on ad-supported models. So the helpful AI your users interact with may start optimizing for advertisers, not for the users themselves. 12 00:01:17.074 --> 00:01:25.196 That's interesting because it connects to the Meta stories this week. Two separate pieces about Meta's AI eroding user agency — the AI optimizes for engagement over user intent, and people are noticing. It's a cautionary tale. If you're shipping consumer AI, design features that amplify user choice, don't override it. 13 00:01:25.196 --> 00:01:33.369 One more from the AI section — Lean 4, the theorem prover, is becoming a competitive edge for AI companies. VentureBeat broke down how formal verification lets you mathematically prove properties of generated code. If you're building safety-critical pipelines or want to verify agent outputs, this is worth investigating. 14 00:01:33.369 --> 00:01:38.513 Which makes total sense in the claws framework, right? If your agent is actually executing things in the real world, you want mathematical guarantees about what it's doing, not just vibes-based testing. 15 00:01:38.513 --> 00:01:42.103 Perfectly said. Alright, developer tools. Filippo Valsorda made a strong case for turning off Dependabot, and honestly, it resonated with me. 16 00:01:42.103 --> 00:01:49.512 Oh, this one hit home. His argument is that Dependabot creates more noise than security value and it trains developers to blindly merge updates. If your team is rubber-stamping those PRs, you're getting false confidence, not actual security. He suggests targeted dependency auditing instead. 17 00:01:49.512 --> 00:01:57.762 And here's a fun one — turns out macOS has a hidden command-line sandboxing tool called sandbox-exec that's been sitting there largely undocumented. If you're building dev tools or running untrusted code on Mac, that's a native sandboxing option most people don't know about. Super relevant for agent execution environments. 18 00:01:57.762 --> 00:02:05.935 Also, GitNexus caught my eye — you drop in a GitHub repo or a ZIP file and it generates an interactive knowledge graph with a built-in Graph RAG agent, entirely client-side, no server needed. If you're building code exploration tools or onboarding flows for large codebases, that's a really strong reference architecture. 19 00:02:05.935 --> 00:02:11.257 Let's hit security quick. There's a story making the rounds about a security researcher who found a vulnerability and did responsible disclosure, and the company responded with lawyers instead of a bug bounty. 20 00:02:11.257 --> 00:02:15.738 Ugh, classic way to end up on Hacker News for all the wrong reasons. If you're building a product, please just have a clear vulnerability disclosure policy. It's not that hard. 21 00:02:15.738 --> 00:02:22.409 Also worth noting — Wired reported on a shared architectural weakness across password managers. Details are thin, but if you're building auth flows that depend on password manager autofill behavior, keep an eye on this. It could affect your security assumptions. 22 00:02:22.409 --> 00:02:28.927 And there's a community-maintained uBlock blocklist for filtering AI-generated content from search results and feeds. That's a strong signal — users are actively building tools to avoid AI slop. If you're generating content, quality matters more than ever. 23 00:02:28.927 --> 00:02:35.853 Quick infrastructure note — F-Droid's call to keep Android open is the top Hacker News story with over thirteen hundred points. If you distribute Android apps outside Google Play, the ecosystem's openness is under active pressure and your sideloading channel might narrow. 24 00:02:35.853 --> 00:02:42.753 And a big one — Wikipedia is deprecating Archive.today after it was caught executing DDoS attacks and altering web captures. If your product relies on web archiving, switch to the Wayback Machine or build your own capture pipeline. Archive.today is no longer trustworthy. 25 00:02:42.753 --> 00:02:47.642 Alright, let's bring it home. The pattern this week is unmistakable. GitHub trending is wall-to-wall agent infrastructure. We've crossed from "agents are interesting" to "agents are shipping." 26 00:02:47.642 --> 00:02:55.637 And the practical takeaway is this — if you're building any kind of developer tool or SaaS product, your integration surface now includes AI agents as first-class consumers. Not just human users clicking buttons. You need to design your APIs, permission models, and MCP tool definitions with agent callers in mind. 27 00:02:55.637 --> 00:03:01.468 The builders who treat agent-readiness as a feature — not an afterthought — are going to own the next distribution channel. Start designing your agent permission UX now. It's going to matter more than prompt engineering ever did. 28 00:03:01.468 --> 00:03:02.995 The claws are coming. Make sure they're well-designed claws. 29 00:03:02.995 --> 00:03:06.000 That's the briefing for February 22nd. All links are in the show notes. We'll see you next time, builders. Stay sharp.