WEBVTT NOTE The Rundown — nextbig.dev daily audio edition, 2026-05-17 1 00:00:00.000 --> 00:00:01.821 Hey everyone, welcome to Builder's Briefing for May 17th, 2026. I'm Alex. 2 00:00:01.821 --> 00:00:07.760 And I'm Sam. We've got a packed one today — the agent integration layer is having a moment, Mitchell Hashimoto dropped a spicy thread about AI psychosis in engineering orgs, and there are some genuinely useful new tools hitting the scene. 3 00:00:07.760 --> 00:00:13.948 Let's jump right in. So the big story today is a project out of Hong Kong University called CLI-Anything. The premise is beautifully simple — wrap any desktop or web application in a command-line interface so AI agents can just operate it natively. 4 00:00:13.948 --> 00:00:20.510 Right, and what's wild is how obvious this feels in hindsight. Like, the Unix philosophy has always been about composable text-in, text-out interfaces. CLI-Anything is basically saying — hey, agents already speak CLI fluently, so let's just make everything a CLI. 5 00:00:20.510 --> 00:00:28.420 Exactly. They've got a registry called CLI-Hub where you can browse pre-built wrappers. So instead of writing custom MCP servers or brittle browser automation for every tool your agent needs, you check the hub first. The integration surface has been the biggest bottleneck in agentic systems, not the LLMs themselves. 6 00:00:28.420 --> 00:00:34.608 And what makes this feel like a real inflection point is that it didn't land alone. Sentry shipped XcodeBuildMCP the same day, AWS dropped their agent-plugins — all three solving the same fundamental problem from different angles. It's convergence. 7 00:00:34.608 --> 00:00:40.596 That's the signal. The agent integration layer is becoming its own category. If you maintain a developer tool and you haven't shipped an MCP interface or a CLI wrapper, you're basically invisible to agentic workflows now. It's table stakes. 8 00:00:40.596 --> 00:00:43.341 Which is a big statement, but honestly, it tracks with what I'm seeing builders actually reach for day to day. 9 00:00:43.341 --> 00:00:48.755 Alright, shifting to AI news. Mitchell Hashimoto — co-founder of HashiCorp — posted a thread that went absolutely viral. Over eleven hundred points on Hacker News. The title? "Entire companies are under AI psychosis." 10 00:00:48.755 --> 00:00:54.843 Yeah, I read this one twice. He's arguing that some companies have just wholesale replaced engineering judgment with blind AI-generated output. Not as a tool in the loop — as the loop. And the failure modes are already showing up in production. 11 00:00:54.843 --> 00:00:59.285 If you're a tech lead, this is your cue to audit where AI-generated code is shipping without meaningful human review. The problems aren't hypothetical anymore, they're piling up. 12 00:00:59.285 --> 00:01:06.496 That's interesting because it creates this weird tension — on the one hand, we're celebrating agent-native tooling, and on the other, we're seeing what happens when you let agents run unsupervised. The answer is obviously human-in-the-loop, but the incentive to skip that step is enormous. 13 00:01:06.496 --> 00:01:13.307 On the performance side, there's a really cool speculative decoding breakthrough called Orthrus-Qwen3. It delivers nearly eight times the tokens per forward pass with identical output distribution. If you're self-hosting Qwen3, this is basically a drop-in speed multiplier. 14 00:01:13.307 --> 00:01:17.774 Eight X throughput without changing outputs? That's not incremental, that's transformational for inference costs. Check the repo for compatible model sizes — link in the briefing. 15 00:01:17.774 --> 00:01:24.311 Also worth flagging — Sean Goedecke put out an analysis showing DeepSeek V4-Flash responds really well to activation steering vectors. You can nudge model behavior without prompting, which could outperform system prompts for tone and style control in production. 16 00:01:24.311 --> 00:01:29.675 Oh, that's a sleeper hit. Steering vectors have been a research curiosity for a while, but having a production-grade model that actually responds well to them? That opens up a whole new control surface for builders. 17 00:01:29.675 --> 00:01:37.760 Alright, developer tools. We already mentioned Sentry's XcodeBuildMCP, but let me give it its own moment. This MCP server gives coding agents direct access to Xcode build, test, and deploy workflows. If you're building iOS apps with Cursor or Claude Code, your agents can now actually compile and validate their own changes. 18 00:01:37.760 --> 00:01:41.877 That was such a frustrating gap before. The agent could write the Swift code all day long, but it had no idea if it actually built. This closes that loop completely. 19 00:01:41.877 --> 00:01:45.844 And AWS dropped official agent-plugins that let AI coding assistants architect, deploy, and manage AWS infrastructure directly. No more copy-pasting from docs. 20 00:01:45.844 --> 00:01:49.512 Between that, CLI-Anything, and the Xcode MCP — three different on-ramps to the same destination in one news cycle. I'd say the trend is confirmed. 21 00:01:49.512 --> 00:01:56.523 One more on the dev tools front — Julia Evans wrote about moving away from Tailwind and learning to structure vanilla CSS. If you're solo or on a small team and those utility classes are adding more cognitive load than they save, she's got a really practical guide to the off-ramp. 22 00:01:56.523 --> 00:02:00.266 I love jvns's writing. And honestly, Tailwind is amazing until it isn't — and knowing when to walk away is a skill. Link in the briefing for that one. 23 00:02:00.266 --> 00:02:06.080 Quick infrastructure note — OpenTofu, the open-source Terraform fork, continues building momentum. If you've been waiting to migrate off HashiCorp's BSL-licensed Terraform, the ecosystem is mature enough now for production workloads. 24 00:02:06.080 --> 00:02:10.621 Which is kind of poetic given we were just talking about Mitchell Hashimoto's AI psychosis thread. HashiCorp's influence is everywhere today — just not always in the way they'd want. 25 00:02:10.621 --> 00:02:16.784 Also, NVIDIA Labs released SANA-WM — a two-point-six billion parameter world model that generates coherent one-minute video at 720p. It's small enough to run on consumer hardware, which makes it interesting for game prototyping and synthetic data. 26 00:02:16.784 --> 00:02:24.818 Quick hits! California is working on a bill that would require patches or refunds when online games shut down. Oracle dropped over a hundred practical database skills guides on GitHub. And there's a delightful deep dive floating around called "You don't know HTML Lists" about semantic markup that's surprisingly humbling. 27 00:02:24.818 --> 00:02:28.411 I also want to shout out the Ploopy Bean — it's an open-source trackpoint pointing stick for any computer. Hardware nerds, you know who you are. 28 00:02:28.411 --> 00:02:35.447 So here's the takeaway. The agent integration layer is crystallizing fast. CLI-Anything, XcodeBuildMCP, and AWS agent-plugins all shipped in the same news cycle, each solving the same problem from different angles — letting agents operate real tools through standardized interfaces. 29 00:02:35.447 --> 00:02:41.935 If you're building agentic products, stop writing custom tool integrations and start consuming these standardized interfaces. And if you maintain a developer tool, shipping an MCP server or CLI wrapper isn't a nice-to-have anymore — it's how you stay relevant. 30 00:02:41.935 --> 00:02:43.931 That's the briefing for May 17th. All the links are in the show notes. I'm Alex. 31 00:02:43.931 --> 00:02:47.000 And I'm Sam. Go build something cool — and maybe let your agent build it too, just... keep an eye on it. See you next time!