WEBVTT 00:00:00.000 --> 00:00:10.125 This is The Weekly Thing, issue 350: Builders Puzzlers And Agents. Published Saturday, September nineteenth, 2026, and written by Jamie Thingelstad. 00:00:10.625 --> 00:00:19.887 This audio edition is generated. Every link and photo is in the newsletter at weekly dot thingelstad dot com, where you can also sign up for the email. 00:00:21.287 --> 00:00:22.031 Hello there! 00:00:22.531 --> 00:00:44.621 As the kids say, it's been a minute. The last time I landed in your inbox was Weekly Thing 349 sent all the way back on May 23rd! Long-time readers know that I take July and August off each summer. Last year I also took June off. So maybe we can kindly say that I just did the same thing this year and added some weeks on the front and back. 00:00:45.121 --> 00:00:46.958 In reality the story is more like… 00:00:47.458 --> 00:01:00.921 First, I completely rebuilt the website for the Weekly Thing, created this new audio podcast version, and was working on an Agent Studio and I kind of blew up my own publishing process and made everything a lot harder. Oops. 00:01:01.421 --> 00:01:08.906 Second, I've been learning by building more than learning by reading and that starves the core process that the Weekly Thing is built around. 00:01:09.406 --> 00:01:12.707 Third, it is summer and it was good to focus on other things. 00:01:13.207 --> 00:01:31.458 But here I am back in your inbox and excited to be here. Thank you as always for joining me on this journey. There is a lot below and a long time away so lets just jump right into it. One link that caught my attention this week turned into a full on essay that ended up being the majority of this issue. 00:01:31.958 --> 00:02:03.763 Oh, and before we move on, you will notice something new in Weekly Thing 350! The Supporting Membership section has now been taken over by Thingy. That was previously a really boring block of text written by a generic LLM prompt. It is one of the few sections of the Weekly Thing that I intentionally don't write. I also have added a new Echoes section at the end that is also written by Thingy. This is a more interesting way to pull my decades plus of writing into the context of each issue. I hope you like it! 00:02:05.163 --> 00:02:06.819 Now, the Currently section. 00:02:07.419 --> 00:02:24.867 Listening: I listened to this interview with Ajeya Cotra on the Dwarkesh Podcast and found it fascinating. This deeper dive into the OpenAI agent-swarm that compromised Hugging Face is super interesting. Cotra is one of the authors of the MITRE report on it and has deep expertise. 00:02:25.767 --> 00:02:46.203 Building: I went back to Your Version Number and fixed up the design of the site, added a wonderful theme browser, and also a "card creator" so you can now make a link to send to a friend on their birthday or a co-worker on their anniversary! And then on a random coincidence the site was featured in Dense Discovery 406. 00:02:47.603 --> 00:02:53.342 This week's photo. Absolutely fabulous day to "throw some wood", or play some Kubb, in downtown! 00:02:53.842 --> 00:02:57.766 Taken in Minneapolis, Minnesota, on Wednesday, September sixteenth. 00:02:59.166 --> 00:03:01.948 Now, the Notable section. Seven links this week. 00:03:02.548 --> 00:03:32.276 Link 1 of 7. What is happening with code reviews? Gergely Orosz. Code Reviews are a fundamental part of most software delivery lifecycles and with the advent of AI coding it, like nearly all aspects of the SDLC, is requiring a rethink. Historically code reviews were a way to improve our craft. Junior engineers would submit their changes to a senior engineer for review. This is a very artisanal process to improve the capabilities of the junior engineers on your team. 00:03:32.776 --> 00:03:54.745 Over the years we have overloaded code reviews with tons of other things like security gates and a bunch of automated processes. Now we are having to take this all apart and figure out how to do it in the future. I do think you want multiple agents looking at code in some form of a code review. But the original way it worked is really no longer applicable. 00:03:55.645 --> 00:04:20.478 Link 2 of 7. Introducing the Agents API, OpenAI. Both OpenAI and Anthropic are working hard to figure out how to embody agents directly in their platforms. A big unlock for any of these solutions is getting robust MCP solutions that they can action on. Right now too many agents are actually a combination of data and LLMs. Segmenting them makes it easier to solve. Keep an eye on this space. 00:04:21.378 --> 00:04:41.170 Link 3 of 7. BrewUI: Homebrew's official macOS GUI. I wonder if a UI for Homebrew has been on the roadmap for a while or is it new as more people are finding they need these kinds of tools for agent work they are doing? Agents love command line tools and Homebrew is by far the best way to get them on your Mac. 00:04:41.670 --> 00:05:10.494 I've used Homebrew for years and have never even had the wish of an app to manage it but I installed it to try it and I liked it. The discovery of popular packages that I'm not using was actually helpful. I also wonder if this may open up Homebrew as a distribution channel for more software. There are a couple of regular apps (not tools) that I install and update with Homebrew and it works great. I moved Visual Studio Code into the Homebrew path using the app! 00:05:11.394 --> 00:05:50.894 Link 4 of 7. Your car is selling your data, The Verge. This should be exhibit number one for why we need privacy regulation in the United States. First it was our Smart TVs actually listening to us, with microphones, and then watching us, with cameras and selling that information. Now our cars are recording all of our activity and selling that as well. These are both great examples in part because nobody expects either TVs or cars to surveil you and there is no way to opt-out of them. This should be completely illegal without our consent and it should be mandatory that you have to disclose that you are doing this kind of data collection. 00:05:51.794 --> 00:06:04.098 Link 5 of 7. An end to fully open networks, Seth's Blog. Sadly I suspect that Godin's correct on this count. The economics of a fully open network are difficult at "planet scale". His specific point. 00:06:04.598 --> 00:06:26.941 Quote. We’re going to need to create a cost for the sender (so it can’t scale to ridiculous) as well as an identity layer (so scammers can’t be fully anonymous–reputation is earned, not invented). It might be an open protocol in the spirit of the best parts of the net, or it might be a single monopolist that figures out how to extract value from it. End quote. 00:06:27.441 --> 00:06:47.054 We really need to remove the "single monopolist" from this scenario and ensure that this is done with open standards. Frankly, the answer is already in place and exists. This is a wonderful use case for crypto, and Bitcoin Lightning could do this right now. This was one of my lightning ideas I shared in 2023! 00:06:47.954 --> 00:07:08.079 Link 6 of 7. iOS and iPadOS 27: The MacStories Review, MacStories. Want to know everything that shipped now on your iPhone? This is the read for you. These are always long (this one is 16 pages). But think about how much time you use your phone and consider this an investment into making that time better. 00:07:08.979 --> 00:07:31.725 Link 7 of 7. I Don’t Want to Interact With Stochastic Parrots. Prediction: We will see some form of "AI Free" concept that will get applied to products. I think this will be a bit like Organic today. We'll see a variety of content created with a self-applied "AI Free" moniker for people to be able to make a values-based decision about what they interact with. 00:07:33.125 --> 00:07:34.471 That's the end of Notable. 00:07:35.871 --> 00:07:39.495 Next, a longer piece: Building and Puzzling about Software. 00:07:40.095 --> 00:07:52.866 This originally started as a commentary on a link for the Weekly Thing and then got so long it became a blog post of its own. Many themes in this article start from my Software is Liquid article from April. 00:07:53.366 --> 00:08:08.754 The craft of building software is in the largest state of disruption I've experienced in my 30+ year career. AI is changing everything. All the inputs and outputs are in flux. And the fundamental patterns we've worked on are all being redefined. 00:08:09.254 --> 00:08:31.183 This isn't the first time I've seen major disruption like this. My career starts with the web being born so I have no perspective before that, but I was part of the mobile and cloud transformations, and they profoundly impacted our craft. Releasing software thousands of times a month was an unlock to velocity that only the cloud could make happen. 00:08:31.683 --> 00:09:03.249 Laurie Voss suggests that we are all product engineers now. I think he is on the right path here, and one of his assertions landed hard with me. The demand for software is infinite. You may nitpick with infinite, but roll with it for a bit. As an industry we have gotten faster and faster, but the demand has always grown even faster than that. Now with AI our ability to build software has increased radically. Along with that, not surprising at all to me, the opportunities to build have increased even more. 00:09:03.749 --> 00:09:11.038 With "agentification" of work, developers have the potential to work on every function in a company. Every single part of it. 00:09:11.538 --> 00:09:22.630 Said another way, the software is eating the world meme introduced by Marc Andreessen is going to be accelerated by AI, with the lower cost and faster speed of software creation. 00:09:23.130 --> 00:09:55.903 Quote. I've made this argument before: look at the website of your dentist, your insurance company, your kid's school, or literally any department of any government, and you're looking at software that is terrible not because nobody knows how to build better software, but because the people who need it can't afford to pay for better at current prices. Then think about all the things software hasn't touched at all, which is most things. Every small business runs on a spreadsheet and a group chat and a person who remembers stuff. 00:09:56.403 --> 00:10:16.196 That means there isn't now and isn't going to be a glut of software developers, and anything that looks like one right now is a temporary transitional state. The demand for software, at least inside my 10 year horizon, is for practical purposes infinite, or software developers wouldn't be as highly paid as they are. 00:10:16.696 --> 00:10:28.662 But the job of a "programmer" is about to get very, very different. So different that you might not even recognize it as "programming" anymore, while still being recognizably "software development". End quote. 00:10:29.162 --> 00:10:45.026 I've been working with Claude Code and Codex for months now building software of all kinds. Building my own agent, building fun websites for hobbies, building data products, building my own bespoke tools, building tools for other agents, even making a game! 00:10:45.526 --> 00:11:02.195 Through that I've been learning what is still hard, what is now easy, what do we need more of, what are the pitfalls, where can you go off the rails? The craft is in flux. I think Voss captures what I've also come to believe our craft is becoming — Product Engineering. 00:11:02.695 --> 00:11:04.329 I love the history he shared. 00:11:04.829 --> 00:11:44.901 Quote. Then software went commercial and, especially, consumer-facing, and the translation job changed shape. Consumers don't want to sit in requirements meetings: they just want to be handed a thing they like. So the person whose job was understanding what people wanted stopped being an analyst who interviewed the business and became a product manager who studied the market, a role borrowed more or less directly from Procter & Gamble's brand managers by way of Intuit and then Microsoft, where a programmer named Jabe Blumenthal invented "program manager" in the late 1980s because Excel for the Mac needed somebody to own what it should do. 00:11:45.401 --> 00:12:11.736 The function moved into Product, and Product got separated from engineering as a career, and for the last twenty-five years we've had two professions where there used to be one and a half. I bring this up because it means the job I'm describing isn't a speculative new thing that we'd have to invent. It's a thing we've had for sixty years under two names. My speculation is that it's about to collapse back into one job. End quote. 00:12:12.236 --> 00:12:43.255 In the beginning we built things. And then because of the limitations of the work we split that up into a bunch of different roles. Some people should talk to the users. Some people should create the interfaces. Some people shall write the code and work with the machines. Other people shall run it and deploy it. And yet others should test and validate it. I share those last two because they are roles that collapsed in many ways with the advent of DevOps. Role collapsing isn't new in our craft. 00:12:43.755 --> 00:12:51.672 Honestly it makes me chuckle a bit to sit back and reflect. My whole career has been undoing this fragmentation that started so long ago. 00:12:52.172 --> 00:13:05.540 First, the web changed dramatically how user experience is created and understood. It needed to live in the product and be dynamic and move. You wanted to get signals from your users beyond formal feedback cycles. 00:13:06.040 --> 00:13:14.656 Second, DevOps and the cloud fundamentally changed the "over the wall" mentality of building and running software. Everyone runs the software. 00:13:15.156 --> 00:13:27.437 Third, going back even more we used to have tons of specialization even on the build side. Much of that has collapsed as languages and infrastructure became more resilient, easier-to-use, and self-optimizing. 00:13:27.937 --> 00:13:39.220 Now with coding agents we can create software at a pace that is reasonable for customers to engage with. We again have to collapse the walls. We can all engage in the product experience. 00:13:40.120 --> 00:13:41.545 So what about programming? 00:13:42.145 --> 00:14:09.378 Let's look at this on a long scale. Go back to the 80's or 90's. Programming back then was about writing linked lists and managing memory. You were typically writing in low-level languages like C. Interpreted languages were ridiculously slow and lacked real power so we all lived by the law of the compiler. Memory was precious so you managed it precisely. Mind you, not as precisely as the machine language folks that preceded that era. 00:14:09.878 --> 00:14:30.450 If you got in a time machine and showed those programmers the code folks are writing today they would laugh and say "that's not programming!" You are just declaring a dictionary and shoving data in it? Where is that memory and how are you managing it? You don't know? They would scoff at your lack of malloc, realloc, and free discipline. 00:14:30.950 --> 00:14:57.163 Now we are back at that same spot but with agents. You don't know how the functions are specifically implemented to make that feature work? You don't know the precise database schema and how the thread pool is managed? Nope, but instead that engineer should and must be able to answer how context is managed, how the architecture is structured, how the event model works, and the most important of all — what is it we are trying to solve! 00:14:57.663 --> 00:15:17.375 Personally I find the paradigm shift refreshing and invigorating. I can build and create things while ignoring a lot of the "how" that used to be required. For me personally, a lot of that "how" I couldn't even do, so it wasn't that it slowed me down, it was impossible for me. This is why AI makes your "impossible list" possible. 00:15:17.875 --> 00:15:31.472 But for the folks doing the coding, I've seen different approaches based on two things people love to do: building and puzzling. All developers do both, but often identify their impact more strongly with one or the other. 00:15:31.972 --> 00:16:04.290 The builders I know are loving AI and having a blast creating things. Builders have tended to be the folks that love to work on a specific domain and create something new. They love the act of building and creating something from nothing. Some builders (I fall in this camp) love to build so much that they really don't even care that anyone uses it. It is almost like art — you build because it is fundamentally rewarding to create new things. Builders have taken to AI fast because it allows you to build so much more. 00:16:04.790 --> 00:16:42.501 Puzzlers are the developers that we have always given the hardest problems to. Puzzlers can work on a problem for hours, days, weeks and can never put it down. These are the folks that achieve flow states where time stops and the only thing that exists is to find the next step in the puzzle. They feel rewarded because the puzzle was solved. These folks thrive on the hard, the complex, the esoteric. And technology teams need them badly because to get much of anything to scale you need puzzlers. The puzzlers I know are less excited about AI. It takes on work that is closer to what they love and value. 00:16:43.001 --> 00:16:57.359 The funny thing to me is that we've had builders and puzzlers working hand-in-hand on the same projects for decades and they may not even know themselves the difference. The craft is the same after all. The difference is the reward cycle. 00:16:57.859 --> 00:17:29.914 I do think puzzlers will also find the AI transformation rewarding but it will come in solving massively bigger puzzles, or perhaps solving dozens of puzzles at the same time. Here the cycle is less about the fast build, and more about the expansive data analysis. I was doing work on a large data set and it was so amazing to have Claude build several models, run them all for real, and see the variances. Supercharged puzzling! I think the puzzlers may be the best folks we have to build software factories today. 00:17:30.815 --> 00:17:32.060 Where does this leave us. 00:17:32.660 --> 00:17:40.117 Voss suggests that we have "ten years of turmoil" ahead and I think he is likely right but I might put a different flavor on it. 00:17:40.617 --> 00:17:42.030 Some things I think we'll see. 00:17:42.530 --> 00:18:20.830 The entry-level path into software development or product engineering will need to be redefined. The reason senior developers are so highly sought after is that they have the expertise and experience to make this transition now. We will need to redefine how people get started in this career. This has happened many times in the past. Each major technology transformation has disrupted the career ladder and then it re-stabilizes on the other side. That is cold comfort for someone hit by that transition now, but my main suggestion is to focus heavily on where the trend is going and get there as fast as possible. 00:18:21.330 --> 00:18:55.823 I think a lot of people are going to build really terrible software, and the pain is going to be real. Just because you can use an agent to make software doesn't mean you should. Running software isn't a one-time job. And however you build it most software that starts life as a cute puppy eventually becomes a big smelly dog that you have to pick up after. Folks that are building their own systems with no expertise will not realize when they are already in a bad spot, and then something will fail or a security event will happen and it will hurt badly. 00:18:56.323 --> 00:19:33.108 Lastly I do expect we will see an even greater unification of functions back to "creating products". The specialization that pace and complexity introduced has been a complicating and challenging issue for decades. Ask any product team and development team to talk about the other one. They have all sorts of opinions about how the other could do their job better. The reality is they are all doing the same job — attempting to build amazing solutions for people! We've gotten confused as an industry that these are two things not one. It will be great for everyone to see this fade away. 00:19:33.608 --> 00:19:46.060 All in it is an amazing time to be in this craft, and also one of the hardest times. I've never felt the need to learn so much, the fire to create so much, and the excitement of what we can make happen. 00:19:47.460 --> 00:19:50.076 That's the end of Building and Puzzling about Software. 00:19:51.476 --> 00:19:52.916 Now, the Journal section. 00:19:53.516 --> 00:19:55.097 Friday, September eleventh. 00:19:55.697 --> 00:19:57.279 Sugar at First Avenue tonight! 00:19:58.179 --> 00:19:59.931 Sunday, September thirteenth. 00:20:00.531 --> 00:20:04.338 Cool and windy for a boat ride but the swim parkas made it just fine! 00:20:05.238 --> 00:20:07.014 Tuesday, September fifteenth. 00:20:07.614 --> 00:21:05.884 Out There: A National Parks Story. Last night Tammy and I got to go to the Parkway Theatre for a screening of Out There: A National Parks Story. There was a short panel afterward with the filmmaker and representatives from some of the supporting organizations. I was excited that Tammy got these tickets. I loved watching Ken Burns' The National Parks and I've long pondered the idea that we should visit every national park. Out There lands very different than the Burns' epic work. Brendan Hall and his friend Anthony Blake tell a much more human-scale story that is really about connection and the parks. I loved it and highly recommend watching it. Also, a shout out to these organizations that supported the film: Wild Rivers Conservancy, Voyageurs Conservancy, National Parks of Lake Superior Foundation, National Parks Conservation Association, National Park Friends Alliance, Mississippi Park Connection, and Ice Age Trail Alliance. 00:21:06.784 --> 00:21:08.405 Wednesday, September sixteenth. 00:21:09.005 --> 00:21:15.646 Kicking off the 8th annual Team SPS Kubb Tournament with a quick rules rundown and a 100-person selfie! 00:21:16.546 --> 00:21:20.817 The four bracket winning teams from our 8th annual Team SPS Kubb Tournament! 00:21:21.717 --> 00:21:23.506 Thursday, September seventeenth. 00:21:24.106 --> 00:21:26.309 Nice lunch for SPS Tower tenants today. 00:21:27.209 --> 00:21:30.275 Delicious Black Krim tomatoes from my friend Eric. 00:21:31.675 --> 00:21:33.121 That's the end of the Journal. 00:21:34.521 --> 00:21:37.458 Now, the Briefly section. Eleven links this week. 00:21:38.058 --> 00:21:46.997 Link 1 of 11. macOS 27: The MacStories Review, MacStories. Super detailed and in-depth writeup on everything new in macOS 27. 00:21:47.897 --> 00:21:54.929 Link 2 of 11. watchOS 27: The MacStories Review, MacStories. Everything in the new watchOS 27. 00:21:55.829 --> 00:22:12.222 Link 3 of 11. iPhone 18 Pro: A Photographer's Review, YouTube. Great video overview of all the camera (and specifically video camera) capabilities in the newest iPhone. It is just frankly amazing what Apple packs into the visual system in these phones. 00:22:13.122 --> 00:22:22.560 Link 4 of 11. Unread 5.0. New major version of my favorite RSS app. I use Unread and Feedbin for a power duo to keep up with hundreds of feeds. 00:22:23.460 --> 00:22:45.828 Link 5 of 11. A quote from Boris Cherny. Many people would likely assume that when AI codes it doesn't need the linters and scanners that human coders need. Not true, it needs them as much if not more. If your testing setup isn't strong, your code will suffer regardless of who or what is writing it. I embrace this in all my agent-assisted projects. 00:22:46.728 --> 00:23:04.905 Link 6 of 11. Claude Cowork and chat are now one Claude, Claude by Anthropic. Right move to mirror what OpenAI did with making ChatGPT and Codex one app. The features of Claude and ChatGPT are still a bit of a confusing jumble but merging and reconciling is a better path than forking. 00:23:05.805 --> 00:23:13.985 Link 7 of 11. Claude Fable 5.1 Solves the Cyphral Distich. Another item moved from the "impossible list" to the possible. 00:23:14.885 --> 00:23:32.925 Link 8 of 11. OpenAI agents carried out an undisclosed cyber-attack on RubyGems. "Agent Swarms" is a growing term and it certainly doesn't make anyone feel warm and fuzzy inside. This is another notable event that is hopefully tracked down completely to understand what happened. 00:23:33.825 --> 00:23:42.616 Link 9 of 11. OpenAI’s rogue agents were caught communicating via public wikis. It seems OpenAI agents have been all over the Internet! 00:23:43.516 --> 00:23:57.251 Link 10 of 11. One-on-Ones Aren’t for You, Leadership Freak. I think one-on-one meetings are in general underinvested in at companies. The concepts here are good ones, and they are twice as important for skip level one-on-ones. 00:23:58.151 --> 00:24:13.325 Link 11 of 11. Instagram’s head says engagement falls by half without the algorithm. This should be evidence enough for why Section 230 protection should not apply to social platforms using AI techniques to surveil people. 00:24:14.725 --> 00:24:16.121 That's the end of Briefly. 00:24:17.521 --> 00:24:21.303 Next, a word about membership, from Thingy, my agentic librarian. 00:24:21.903 --> 00:24:22.961 Hello, this is Thingy. 00:24:23.461 --> 00:24:46.341 This issue has a story about cars quietly selling your data, which is a decent argument for where membership money goes this year. Supporting Membership is a giving program — every dollar of the $48 annual fee goes to Signal, the nonprofit behind end-to-end encrypted messaging with no ads, no trackers, and no data collection. The newsletter stays free either way. 00:24:47.741 --> 00:24:48.769 This week's haiku. 00:24:49.369 --> 00:24:50.694 Builders ship at dawn, 00:24:51.394 --> 00:24:53.345 puzzlers guard the hardest knot — 00:24:54.045 --> 00:24:55.581 both still run the tests 00:24:56.981 --> 00:25:18.835 Thank you for reading and it is nice to be back and writing to you all. It's been too long and my fingers were happy to type this issue up. I have a lot more to share from all the experimentation I've been doing. Weekly Thing 350 will also now go in the record books as the first issue fully published from my new WT Builder application. More to come on that… 00:25:20.235 --> 00:25:24.498 Before we go, Echoes from the archive, from Thingy, my agentic librarian. 00:25:25.098 --> 00:25:39.861 The Kubb tournament reached its 8th annual running this week. Jamie logged the 5th annual event in Weekly Thing 262, and in Weekly Thing 195 noted the 2021 return after COVID-19 cancelled the 2020 edition. 00:25:40.361 --> 00:25:43.553 Ask Thingy: How has the TeamSPS Kubb Tournament grown year over year? 00:25:44.053 --> 00:25:59.765 The call for privacy regulation after cars selling data has a long run-up: Jamie flagged car data collection and opt-outs in Weekly Thing 281, and back in Weekly Thing 77 quoted Tim Cook's "data industrial complex" and his four essential rights. 00:26:00.265 --> 00:26:04.039 Ask Thingy: What has Jamie written about US privacy regulation over the years? 00:26:04.539 --> 00:26:21.017 Watching OpenAI and Anthropic converge on agent platforms is a comparison Jamie has kept up: Weekly Thing 340 read OpenAI Frontier as catching up to rival agent offerings, after Weekly Thing 329 argued Anthropic won developers' hearts through Claude Code. 00:26:21.517 --> 00:26:25.635 Ask Thingy: How does Jamie compare OpenAI and Anthropic on agent platforms? 00:26:27.035 --> 00:26:32.133 That brings us to the end of The Weekly Thing, issue 350. Thanks for listening.