Direct link to podcast audio file

The promise of Breaking Change is that with every major version (this being the 60th episode of the program, but only the 53rd such release), I will break something. Well, I finally did it. I think I broke the show. Find out how by listening for yourself!

Want to be my friend? Drop me a line at justin@searls.co. Want to file a complaint with my supervisor? Forward your concerns to podcast@searls.co.

These links may rot, but alas—death comes for us all:

• Tonal (without a subscription)
• Better with Becky
• My appearance on the freeCodeCamp podcast
• Dual-loop BDD is the new Red-green TDD
• Models are commodities, harnesses are differentiators
• Aaron's puns, ranked
• United Airlines can now boot passengers who refuse to use headphones with their devices (News+)
• Your Meta glasses are watching you pee (9to5mac's summary)
• Tuna for Mac by friend of the show, Mikkel Malmberg
• MacBook Neo
• MacBook Neo offers more single-core performance than any mobile processor from AMD, Intel or Qualcomm
• Apple + Nvidia's CloudXR & Foveated streaming partnership: 1, 2, 3, 4, 5
• Xbox's Project Helix
• Microsoft's 'Xbox mode' is coming to every Windows 11 PC (News+)
• Windows: Announcing new tools and platform updates for Windows PC game developers
• Sony is testing dynamic pricing in the PlayStation Store
• DLSS 5 should make you excited and/or outraged: video, backlash
• Gamblers trying to win a bet on Polymarket are vowing to kill me if I don't rewrite an Iran missile story
• South Korean Police Lose Seized Crypto By Posting Password Online
• China mobilizes "one-person company" AI startups
• Scrubs is back!
• Mickey 17 is not about a cartoon mouse
• The Expanse
• Marathon - and how sickos like me can't stop thinking about it

This morning, Zed (a native Rust app) was 500% CPU without a single window open, 3 macOS internal daemons had crashed, 800ms input lag.

• Running Claude Code in three tabs at once
• Compiling multi-package Xcode builds
• Automating two iOS Simulators and a Mac build of my app

Hasn't missed a beat. Not bad for a phone chip.

Maybe Anthropic is just doing humanity a solid by helping us understand how much it will suck when the era of subsidized pricing for LLM-based products ends.

Direct link to podcast audio file

Quincy Larson over at freeCodeCamp had me on their podcast to discuss how the rapidly changing software industry is impacting junior developers and what they can do about it. I don't normally spend time talking about this stuff, because I started programming in the 90s and can't claim to know anything about what it's like to just be starting now. I've also always discouraged people from getting into software development unless they're super passionate about it, which has been out of step with the "learn to code" hype train that gathered steam over the last fifteen years only to run into a brick wall recently.

His audience is way larger and composed of quite different people than those who follow my stuff, so I strongly recommend you read every last YouTube comment. Some pretty crazy shit, if I'm being honest. Makes me glad I don't have 11 million subscribers.

Appearing on: freeCodeCamp Podcast
Published on: 2026-03-17
Original URL: https://www.freecodecamp.org/news/there-are-2-kinds-of-devs-one-of-them-is-screwed-justin-searls-interview-podcast-210/

Comments? Questions? Suggestion of a podcast I should guest on? podcast@searls.co

It follows, then, that if agents require integration between model and harness, that the companies building that integration—specifically Anthropic and OpenAI—are actually poised to be significantly more profitable than it might have seemed as recently as late last year.

This assumption is incorrect. To date, agents have not clearly benefited from proprietary integration to their favored model. I thought this was obvious, but I sometimes forget that not everybody else has made it their full-time hobby to mix and match various coding agents with various large language models for the past year.

Here's how I see it.

The API surface of a frontier model is text in, text out. Images in, images out. Audio in, audio out. AI models are among the most immediately pluggable and therefore commoditizable innovations in the history of software—right up there with UNIX pipes and their simple promise of text in, text out. Countless forks and TUIs have already demonstrated that swapping the underlying model is a marginal concern, not a structural one. The harness doesn't integrate with the model. It integrates with the user's world—files, tools, workflows, intent—and calls the model as one resource among many.

This distinction matters enormously. A high-quality harness paired with a mediocre model can accomplish incredible things. A frontier model paired with a poorly-considered harness can barely get out of the gate. If model-harness integration were truly the differentiator, the model would indeed be irreplaceable and invaluable. But in practice, models have turned out to be the most replaceable part of the stack. And every recent advance in tool use, verification, and orchestration has only made this clearer—the harness gets more valuable while the specific model behind it gets more interchangeable. As long as multiple companies are credibly in the frontier model game, there will be cutthroat competition and relentless downward pressure on price. And while it may not happen soon, models and hardware will eventually advance to the point that a local open model is good enough—and then the frontier labs won't just be commoditized, they'll be cut out entirely.

Thompson brings up Microsoft's new Copilot Cowork bundling initiative as evidence that model-harness integration is where value accrues:

Fast forward to last week, however, when Microsoft revealed how they will handle the potential business impact of AI reducing seats, which is a bit of a problem for their seat-based business model: the company is going to bundle AI into a new higher-tiered enterprise offering, E7, which is going to cost twice as much — $99 per seat per month — as the formerly top-of-the-line E5. That's a big increase, which Microsoft needs to justify with AI that actually makes those seats more productive, and the product they launched with the new bundle was Copilot Cowork.

Microsoft is indeed in a tight spot, but not because the model makers have a monopoly on harnesses. Instead, blame their outlandish investments in somebody else's model, their piss-poor track record for delivering a single worthwhile Copilot-branded piece of software outside the GitHub org, and their over-reliance on per-seat licensing that is existentially at odds with a strategy that, if successful, would drastically reduce the number of seats its customers need. None of that tells us anything about whether models and harnesses are structurally coupled. It only tells us what we already know: that Microsoft has put itself in a strategically weak position—a strategy that presupposed they would somehow stop sucking at building software people actually want to use.

Apple's situation is a cleaner test of the thesis. Dediu is probably right that Apple has a tremendous opportunity precisely because they can build a harness and swap the underlying model. Will they build an incredible harness? No—they'll build one that's just good enough to keep customers buying their hardware. But the strategic logic is sound: focus on the human interface and the real-world problems it can solve, and don't cede the narrative to implementation details like which model is running underneath. The truly great harnesses will probably come from smaller, more nimble startups who understand that the opportunity isn't in training the next frontier model—it's in connecting the horsepower of any capable model to the thousand different domains waiting to be served.

The frontier labs attracted an outsized share of attention and funding because OpenAI happened to hit one out of the park with ChatGPT as a product, and this conflation continues to produce analyses that make the same mistake Thompson is making. ChatGPT went so mainstream that it's treated as a blockbuster product, but I suspect we'll look back on it merely as a (historically impactful) proof of concept. The work of building a chatbot and the work of training a frontier model are only incidentally related. The thousands of chatbot systems that have sprung up since ChatGPT's release should be all the evidence we need that building a great chatbot never required building a great model—OpenAI just happened to do both.

At the end of the day, the utility of models is owed to the fact that they are black boxes with simple interfaces. Text in, text out. Audio in, audio out. Increasingly, video in, video out. But that simplicity is also what makes them commodities. And leads don't last long enough for anyone to establish a moat—the underlying research unlocks improvements in the open, and ubiquitous access to everyone's models facilitates distillation of competing ones.

Harnesses couldn't be more different. Harnesses are the gooey, ever-shifting integration of human-computer interface. They connect people's needs to implementations designed to meet them. With a little cleverness and elbow grease, they can be dispatched to verify that those implementations work. They can be assigned to orchestrate other agents across systems in separate geographies, and over sessions conducted at different times.

The model is the engine. The harness is the car. And history has shown—over and over—that you don't need to manufacture your own engine to build a great car.

🔗 stratechery.com

🧂 justin.searls.co

I fucking suck, but even playing solo it can feel incredible when you do manage to come out on top. If you want to roll with me, I'm Searls#2430 (or just email me justin@searls.co)

For the last month, I've been working on prove_it, a framework for building quality harnesses for Claude Code—primarily via its hooks system. In a recent release, I added TDD enforcement to its default configuration. First, it injects a test-first development approach into every plan Claude generates. Then, a PreToolUse hook follows up with permissionDecisionReason reminders whenever the agent deviates from the one true path (e.g., repeatedly edits source files without touching any tests, never runs a test to see it fail, etc.).

And just like real life, nagging developers works. My current side project is having no problem maintaining 100% (not 99%, not 99.9%, but 100%) code coverage.

My initial prompt simply told the agent to practice "red-green TDD" (a phrase I had never heard of until it was discovered that LLMs apparently interpret it as "real TDD"). This approach turned out to be woefully insufficient. Why? Because agents follow the path of least resistance and will invariably write a shitload of unit tests chasing the local maximum of code coverage without any regard for the global maximum of making sure shit actually works. After finishing each feature, all the code would have real unit tests with real assertions, but each time I tried running the fully-integrated app, the agent would find a novel way to miss the forest for the trees.

So yesterday I updated the prompt with the more sophisticated dual-loop approach developed by folks like Dan North and other adherents of behavior-driven development in the late aughts. It is best illustrated by two concentric circles: you begin each feature with a failing integration test, then dive into an inner loop for numerous red-green-refactor iterations of unit tests, then pop back out again once the outer loop's integration test passes.

Honestly, I hadn't taken this practice off the shelf in a while, but since updating the prompt that gets injected into every Claude Code plan, it's been working out great:

## Development approach

Follow BDD dual-loop TDD. Every feature increment starts from a failing integration
test and is driven inward through unit-level red-green-refactor cycles.

### Outer loop (integration)

1. Red (integration) — Write one integration/acceptance test that describes the
next observable behavior from the outside in. Run it. Confirm it fails for the
reason you expect. Do not proceed until the failure message matches your intent.
2. Inner loop (unit) — repeat until the integration test can pass:
  - Red — Write the smallest unit test that expresses the next missing piece of
implementation the integration test needs.
  - Green — Write the minimum production code to make that unit test pass.
Run it in isolation and confirm. No speculative code.
  - Refactor — Clean up the code you just wrote (duplication, naming, structure)
while all unit tests stay green. Only touch code covered by passing tests.
3. Green (integration) — When enough unit-level pieces exist, re-run the
integration test. If it still fails, diagnose which piece is missing and drop back
into the inner loop. Do not add code without a failing test driving it.
4. Refactor (integration) — With the integration test green, refactor across
module boundaries if needed. All tests — unit and integration — must stay green.
5. Repeat from step 1 with the next slice of behavior until the task is complete.

### Discipline rules

- Never skip the red step. If you cannot articulate why a test fails, you do not yet
understand the requirement.
- One logical change per cycle. If you are changing more than one behavior at a
time,
split it.
- Run only the relevant test after each green step, then the full suite before each
commit-worthy checkpoint.
- If a refactor breaks a test, revert the refactor — do not fix forward.
- Treat a surprise failure (wrong message, wrong location) as information: re-read
it,
 adjust your understanding, then proceed.

I imagine most people reading this will just copy-paste it into an AGENTS.md file or something, but if you're actually interested in learning more about this topic, my favorite articulation of this concept can be found in the edition of The RSpec Book that Zach Dennis worked on.

Now for just a touch of resentment

Not for nothing, but a friend of mine asked why it was that so many programmers who had previously rejected test-driven development and related practices are suddenly embracing it. I genuinely believe some people interpret the suggestion that their code isn't good enough as a personal affront. They feel that being told to write tests, much less orient their workflow around verifying the quality of their work product, is somehow an indictment of their programming skills. So fragile is the ego of many programmers. I witnessed this defensive reaction firsthand on countless training and coaching engagements, so I'm speaking from experience here.

Of course, in 2026, the same people are suddenly huge fans of the same practices they once dismissed, because now we're talking about verifying some dipshit AI agent's work. The key difference is that any such tests exist not as a condemnation of themselves, but of the code written by some external thing.

What's funny about this, of course, is that nothing has really changed. If you zoom out, it's still just some doofus staring at a computer screen in silence all day. But yeah, a lot more people are suddenly really interested in TDD than there used to be.

I'm still iterating on my experimental Claude Code verification harness, prove_it. This week my focus has been on nudging agents to practice test-driven development. Traditionally, we called this "TDD", but which has recently been renamed to "red-green TDD" as it has been discovered that this is what LLMs interpret as "real TDD".

Anyway, so that I could watch it steer an agent in a fresh codebase in real time, I opened a new directory and asked Claude Code to one-shot a terminal-based tennis game, replete with scoring and an AI that I couldn't beat. In OCaml. And it worked! It actually test-drove everything. Neat!

Direct link to podcast audio file

Today, we're joined by a very special guest, Len Testa! You might know him from The Disney Dish podcast or from his excellent theme park travel planning app Touring Plans. Or you might not know him at all! No wrong answers.

This episode is all about Disney's ill-fated Star Wars: Galactic Starcruiser live action role-playing hotel—which we both had the opportunity to experience right when it first launched. It was a life-changing experience for both of us, so why did it fail? Is it true that the top Disney brass learned all the wrong lessons from that failure? And will the CEO-in-waiting Josh D'Amaro ever have the courage to attempt something so ambitious again? Listen to this episode, in which we speak with an unearned confidence that suggests we have all of these answers!

You can reach out to Len at len@touringplans.com and you can write into the show at podcast@searls.co. He'll read your e-mail and reply to it, but I may only skim it.

You can't pay me to say nice things about you, but you CAN pay me to not say mean things about you. Please send inquiries to sponsorship@searls.co theverge.com/news/889595/tim-sweeney-signed-away-his-right-to-criticize-google-until-2032

Direct link to podcast audio file

Sure feels like some combination of AI, the US military, and the AI military could bring an end to the world any day now, so I figured I'd better record one last show for posterity. Welcome me on this version's speedrun to the apocalypse!

So long as the EMP blasts don't nuke all our ham radios, write in to podcast@searls.co and I'll read it on the next release of the program. Over.

Be sure to click all these links while the clickin's good:

• Ninja Creami Deluxe
• prove_it
• Agents are ushering in the Antisocial Era
• Brace for the Fuckening
• Citrini Research fanfic
• Would you like your AI singularity benign or extinction-level?
  • Original post by the Dallas Fed
• Anthropic's Statement on the comments from Secretary of War Pete Hegseth
• OpenAI's agreement with the Department of War
• U.S. Strikes in Middle East Use Anthropic, Hours After Trump Ban
• Human brain cells on a chip learned to play Doom in a week
• Apple introduces a new video podcast experience on Apple Podcasts
• Foveated Streaming for Apple Vision Pro
• Huge Xbox shakeup: Phil Spencer retiring, Sarah Bond resigns, Microsoft AI boss becomes new Microsoft Gaming CEO
• XCOM's Jake Solomon announces his new studio is closing, reveals now-cancelled AI-driven narrative game
• Fan group makes Unreal Tournament 2004 available to play for free and in 4K
• Video Games Are Losing a Brain Rot Race to Gambling and Porn (News+)
• Netflix walks away from its deal to buy Warner Bros. after Paramount came back with a better offer (News+)
• Why Netflix Actually Won Even Though It Lost Warner Deal (News+)
• Aaron's puns, ranked
• Your Friends & Neighbors
• Rental Family
• Halt and Catch Fire (again)
• Paradise: Season 2
• Star Trek: Voyager - Across the Unknown
• xcbeautify
• turbocommit
• The double loop model for agentic coding

Generally, I set my expectations for sushi in Florida to quite low, especially if the restaurant isn't Japanese-owned. Izuki blew away those admittedly reserved expectations with one of the best sushi courses I've ever had—and it's just around the corner from our place!

Beyond having a direct relationship with a supplier that ensures several shipments a week, what really sets Izuki apart (other than excellent preparation) is their dry aging technique. The true flavor of each fish is bolder and more distinctive, and the textures really melt in your mouth. If you're in Orlando, you've got to try it. One of the few Japanese restaurants in the United States I wouldn't be ashamed to bring my Japanese friends to.

Thanks to the upheaval being caused by coding agents, I believe we've entered a new era without realizing it, one that might be called Antisocial Coding.

Three things I've been chewing on lately:

• Agents blow out team communication costs. Last year, Scott Werner compared a developer orchestrating agents to an octopus with 8 autonomous arms and I replied to point out how this creates a hub-and-spoke communication crisis that would prove unwieldy for traditional multi-developer collaboration. Since then, I've been advising startups to stay single-developer for as long as possible and telling larger organizations to start moving to one-repo-per-human wherever they can—even if it means re-architecting their systems to align with this.

• Agent-built codebases calcify quickly. Last night, Caleb Hearth texted to tell me he'd created several issues on turbocommit, and I'm ashamed to admit my brain was so overwhelmed with six terminal tabs of unrelated WIP features that I brusquely denied all of them. Caleb shared a keen insight: many agent-led codebases are becoming impenetrably large and complex extremely early in their lifespan—even while the userbase is still limited to the original creator. And on reflection, I do sense a certain ossification in the projects I encounter lately—so tightly coupled to one person's hyper-specific needs that they lack room to grow or change by the time other users arrive. Beads is the perfect example: I spent two weeks trying to use it, but at 4 months old it was already far too complex and brittle to integrate into my workflow

• Open source is closing its doors. Yesterday, tldraw took the unusual step of removing the tests from their codebase, on the theory their continued presence would make it far too easy for anyone with an agent to build a cleanroom rewrite and undermine their business. Nothing says "closed for contributions" quite like not even having access to a test suite or CI to work with

I don't feel strongly about any of this, because I'm mostly a loner to begin with. But this trend will likely have troubling impacts for teams and organizations.

• If the last decade of social coding taught us anything, it's that pinning the continuity of your business operations on a bunch of software for which the bus factor is 1 is an unacceptable liability. If the next decade is typified by an explosion of solo projects, then the rational response will be to adopt fewer dependencies and have agents build more in-house implementations—even if it means shouldering more maintenance costs and security risks
• Unlocking the maximum productivity afforded by agents by sequestering individual human developers to their own discrete fiefdoms (whether at the repository, CODEOWNERS file, or some other boundary) will only accelerate the trend of discouraging contributions from others, compounding the above issue at scale and throughout engineering organizations
• While many people are (rightly) concerned about AI's impact on the job prospects of less-experienced developers, the continued erosion of collective code ownership coupled with a trend away from inter-person collaboration will further slam the door on opportunities for mentorship and apprenticeship, exacerbating the widening gap in potential value creation between juniors and seniors

I've never had more fun building stuff than I'm having right now, but that doesn't change any of the above. These are important things to be watching out for. Engineering leaders should probably consider the downstream effects discussed here before plowing ahead with plans to reorient their organizations around what's best for maximizing the productivity of semi-autonomous coding agents.

No wonder my inbox has been so peaceful! github.com/orgs/community/discussions/157470

turbocommit does this by linking each session's commits: github.com/searlsco/turbocommit?tab=readme-ov-file#continuity-across-workstreams

Their customer support contends that it is. I feel pretty strongly that it's not.

Who's to say. apple.news/AaV5OdPkJR12OiHpdMqOdYQ

Answer: if there were a process or playbook to make yourself irreplaceable, AI agents would follow it. And replace you.

iOS 5-26: idiotic waste of time
iOS 27+: productivity life hack

Humans who lack that gear are absolutely fucked.

So FYI, I just had a great fucking idea. Feelin' cute.

Expect the Starbucks in your local suburb to become occupied with middle-aged former office workers who want to get out of the house. That's a benign portrait, but a lot of these families still owe mortgages on their houses that they won't be able to maintain. If I were a homeowner in Silicon Valley or Westchester County, I might consider putting my house up for sale to see what I could get because there's going to be downward pressure on these communities. It might not feel great being first, but you don't want to be last.

Then again, why should I be surprised that politicians aren't thinking about this? Almost nobody is! AI boosters are having too much fun playing with Claude Code to connect the dots between its current capabilities and the post-office-work utopia/dystopia that awaits us. Meanwhile, AI skeptics seem worryingly self-assured in their wishcasting that today's agents:

1. Will never be able to do the things they can already do
2. Will forever remain as incompetent as they are now
3. Will disappear the moment the AI economic bubble bursts

Other than the first item on that list, I don't even know that the skeptics are wrong! Maybe they're right! In fact, it'd be more convenient for all of us if the skeptics end up being right. But are we really certain that the probability Sam Altman is even kinda-sorta right is literally zero percent? If not, then we should probably prepare for that potentiality.

I imagine that we all have a pretty clear image of what will happen if the AI bubble pops: stocks will crash, Steam Decks will be back in stock, and Sam will end up in jail. (Maybe every disastrous hype cycle in tech simply demands that at least one awkward white man named Sam goes to jail.) But I've got to admit, many people I talk to and follow online don't seem to have given a second thought to the sort of economic hellscape we're in for if the astronomical valuations of companies like OpenAI, Anthropic, and SpaceX (lol) turn out to have been appropriately priced.

Well, give it thirty seconds' thought, and—to borrow Yang's term of choice—we'll be in for The Fuckening: a macroeconomically significant decrease in the number of high-paying white collar jobs the market will support.

Tech CEOs are bullshitting us

Listen to Satya or Sam or Jensen talk about this "What if?" scenario, and you'll hear them compare AI to technological revolutions of the past. They each breathlessly declare, "there will be new jobs," and that AI will, "create a bunch of new ones," and even that, "there will be more jobs." But, as someone who's been using coding agents to build good-enough software with the velocity of a traditional engineering team for nearly a year now, none of these reassurances pass the smell test. The idea that there will be at least one net-new job created for every job eliminated by AI simply isn't credible.

Remember, for the purposes of this post, we're playing out the scenario labeled "what if these AI investments actually pan out?" That scenario isn't compatible with a future where companies find themselves needing just as many highly-compensated humans as they do today—the ROI needed to justify this level of investment simply wouldn't be there.

So why are the AI executives talking out of both sides of their mouth? Because, what else the fuck are they supposed to do? Go on the record and say, "yeah man, I dunno, now's probably a good time to learn how to can your own food or sell your house if you happen to live in a mid-market metro with a strong services sector."

Whenever I hear one of these guys' bullshit quotes, I have to mentally sprinkle in the unspoken context to protect my own sanity:

• "There will be new jobs for people with deep domain expertise and unusually strong critical thinking skills."
• "Looking at jobs as a debt counselor or estate auctioneer? AI will create a bunch of new ones."
• "If you can swallow a massive pay cut and don't mind shutting doors for a living, there will be more jobs."

Programmers are first in, last out

Software developers may be among the first to experience what it feels like for AI to do our jobs for us, but I still believe we'll be the ones shutting the lights off on the middle class.

In fact, programmers are about to become as busy as we've ever been. Why? Because custom software has historically been supply-constrained. How many businesses would have been created over the last twenty years if even extremely basic apps didn't require millions of dollars of venture funding? Already, agents can help capture those opportunities on a relative shoestring budget. A tremendous number of niche, overlooked cottage industries are finally ripe for their ✨disruption✨ moment.

And all those projects still need human programmers! Why? Because today's agents are nowhere close to being able to write software that won't fall over without supervision. The last 10% needed to reach full autonomy will take 90% of the effort. My Tesla has driven me everywhere for over a year, but it'll be decades before the world stops manufacturing steering wheels.

Programmers may be spread across more companies and working in smaller teams, but there likely won't be a net reduction in the amount of total work to be done anytime soon. But can we honestly say the same about the accountant who's been rendered redundant? Or the junior lawyer position the firm opted not to fill? Or the management consultant who realizes their client can buy the same peace of mind with a $20 ChatGPT Plus subscription? Where are they all going to go? The answer is obvious: if this whole "AI thing" ends up becoming what it's cracked up to be, those people are well and truly fucked. And sooner than they probably realize.

Make your work worthwhile

Ultimately, unless we see AI crash and burn catastrophically, I feel pretty comfortable predicting that if the only sounds your job makes are *clickety-clack* and *yackety-yack*, the bulk of your colleagues—and very possibly you—are going to be caught up in the gravitational vortex of what Andrew Yang is calling The Fuckening.

So, what can you do to mitigate these risks and protect yourself? All you need to do is follow my fool-proof two-step plan:

1. Acknowledge that the odds we're living in The Fuckening timeline are greater than zero
2. Plan accordingly

If you're a programmer and your current role doesn't resemble that of a full-breadth developer, do something about that ASAP. If you're anyone else, then follow this bit of advice buried inside that larger post:

Figure out how your employer makes money and position your ass directly in-between the corporate bank account and your customers' credit card information.

Do the work: trace your individual contributions to the total revenue and cost savings they represent for your employer. Is that number demonstrably higher than your fully-loaded compensation? If so, you should be okay. And whenever the AI tools in your space improve appreciably, re-examine your contributions and run the numbers again. But as soon as you cost more to your employer than you bring in, don't assume they'll keep you around—figure out how to increase your output! And if the numbers are totally upside down, find someplace new where your cost will be commensurate with your value.

All I have is individualized advice, because our individual situations are all we have control over. Yes, there are countless collective actions we could take as a society to mitigate or even eliminate these downside employment risks. There are probably even public policy prescriptions that could position civilization to absolutely thrive in the age of AI. But the odds of either of those happening in the current political climate are even lower than (ugh) Sam Altman ending up being proven right about all this shit.

I'm expressly not suggesting we treat each other as if we're living out a workplace adaptation of The Hunger Games here. I'm just reminding you to secure your own oxygen mask first before assisting others.

Not saying programmers will be remembered as coal miners or 9/11 heroes or anything, but I'm not not saying that.

Welcome to the identity crisis, fellow office workers!

Direct link to podcast audio file

I feel like I'm getting saltier, and it's concerning. If you want me to tone it down and/or up, let me know. It'll go a lot better if you just write to podcast@searls.co instead of yelling at your phone.

As usual, I brought the goods. Now here are the receipts:

• prove_it is doing its job, mostly
• Aaron's puns, ranked
• Starsand Island studio confused by alleged 'praise-bombing' attack
• Sony is considering holding back PlayStation 6 until 2028 or 2029
• Southwest changes are infuriating fans (News+)
• NYT: All the news that's fit to manspread
• Gurman: Tesla CarPlay Held Back by Need for Wider Adoption of Apple's iOS 26 (Archive)
• YouTube on Vision Pro (News+)
• Apple Announces Special Event in New York, London, and Shanghai on March 4
• Filed my first SwiftUI bug as feedback. It's FB21962656
• Thoughtworks concludes TDD is good and billable juniors valuable
• Ugh, Anthropic CEO was right about the timing of AI writing all the code
• Peter Steinberger Chose OpenAI. The Code Was Never the Point
• Eternity
• Starfleet Academy
• Your Friends & Neighbors

Lotta folks losing the thread here. blog.davemo.com/posts/2026-02-14-deterministic-core-agentic-shell.html

The improvement from v13 to v14 is remarkable. Tap Start and it pulls out of the garage, drives you, and parks itself. Another tap and it drives home, opens garage, parks itself. More advanced than people assume.

Claude Code doing a better job self-verifying

Anyway, that's why all your videos sound like shit. You're holding it wrong.

Example: it spun up a complete example project and ACTUALLY proved my cached test runner works github.com/searlsco/prove_it

I just write prompts between sets.

Let users create symbolic links mapped to HTTP resources. Since you can't literally link a file to a URL without changing the actual file system, hyperbolic would centrally handle journaling, cache/etag, updates.

We ordered a wood chest that arrived with cosmetic damage. After logging the damage in their customer support interface, it prompted me to start a chat with their AI virtual assistant.

What happened next:

1. It immediately offered me a 15% refund to keep the product
2. I asked for 20% and it immediately agreed
3. I asked for 25% and it immediately agreed
4. I asked for 30% and it turned me down
5. I took the 25%, which was, indeed, immediately refunded

Turns out that negotiating with a rules engine is way easier than negotiating with a human tasked with operating a rules engine.

So basically, all Wayfair did was add a chatbot to the end of their existing "Report a Problem" interface that will give customers more money if they ask for more money. What a world. 🌍

Instead of giving up, I forced myself to reproduce all my manual commits with agentic ones. I literally did the work twice. I'd do the work manually, and then I'd fight an agent to produce identical results in terms of quality and function (without it being able to see my manual solution, of course).

This was excruciating, because it got in the way of simply getting things done. But I've been around the block with non-AI tools enough to know that friction is natural, and I can't come to a firm, defensible conclusion without exhausting my efforts.

But, expertise formed. I quickly discovered for myself from first principles what others were already saying, but discovering it myself resulted in a stronger fundamental understanding.

1. Break down sessions into separate clear, actionable tasks. Don't try to "draw the owl" in one mega session.
2. For vague requests, split the work into separate planning vs. execution sessions.
3. If you give an agent a way to verify its work, it more often than not fixes its own mistakes and prevents regressions.

More generally, I also found the edges of what agents -- at the time -- were good at, what they weren't good at, and for the tasks they were good at how to achieve the results I wanted.

I recorded an interview on the freeCodeCamp podcast a few days ago saying the same thing. Namely, that this reminds me of every other time programmers have needed to learn a new way to do something they already know how to do some other way. When I was teaching teams test-driven development, I always had to encourage them to force themselves to test-drive 100% of their code without exceptions so they're forced to actually learn the difference between, "this is hard because it's a bad tool for the job," and, "this is hard because I've not mastered this tool yet."

Over-application of a tool is an important part of learning it. And it applies just about every time we're forced to change our ways, whether switching from Windows to Linux, a graphical IDE to terminal Vim, or from Google Drive to a real filesystem.

Either you're the type who can stomach the discomfort of slowing down to adopt change, or you're not. And many (most?) programmers are not. They're the ones who should be worried right now.

🔗 mitchellh.com

🧂 justin.searls.co

My friend Eric Doggett became a Disney Fine Artist a couple years back and he's currently being featured at EPCOT's 2026 Festival of the Arts. Each day this week, he's holding court to talk to people about his work at a pop-up gallery just outside the Mexico Pavilion. Myself and a few other friends ganged up on him this afternoon to lend our moral and financial support by showing up and buying a few pieces.

I really like the painting I picked up. It's a semi-subtle ode to Big Thunder Mountain, a celebration of Walt's love of trains, a not-so-hidden Mickey-shaped rockface, and a tiny nod to the goat.

If you're a local, swing by and say hi to Eric—he's great! If you're not, check him out as @EricDoggett on YouTube—the videos of how he works are pretty cool. I immediately hung it in my office / studio when I got home, because Eric's audio engineering talents are a big reason why Breaking Change sounds as good as it does!

Direct link to podcast audio file

Elon has combined 3 of his 4 businesses and everything makes sense. Also, I went to Japan and all I came back with was another weird story about animal sperm. Other stuff happened too, but let's be honest, it's the typical AI schlock you've come to expect from this decade.

Write in with your own takes to podcast@searls.co. Please. Really! Do it.

Citations needed?

• 白子揚げ = 🐟🍆💦🍟
• scrapple scrapes Apple docs for agents
• prove_it forces Claude to verify his shit works
• Aaron's puns, ranked
• A sane but extremely bull case on Clawdbot / OpenClaw
• US Insurer 'Lemonade' Cuts Rates 50% for Drivers Using Tesla's 'Full Self-Driving' Software
• Elon Musk's SpaceX acquiring AI startup xAI ahead of potential IPO
• Amazon commingling practices will end effective March 31, 2026
• UPS to cut additional 30,000 jobs in Amazon unwind, turnaround plan
• Microsoft is experimenting with a top menu bar for Windows 11 (News+)
• How DoorDash and Other Food Delivery Apps Are Reshaping Mealtime in the U.S.
• Apple debuts 'Apple Creator Studio' subscription, here's what you get
• Apple Sued by App Developer Over its Continuity Camera
• Halide cofounder Sebastiaan de With joins Apple's design team
• Inside Apple's AI Shake-Up, AI Safari and Plans for New Siri in iOS 26.4, iOS 27 (Archive)
• Apple introduces new AirTag with expanded range and improved findability
• Xcode 26.3 unlocks the power of agentic coding
• Josh D'Amaro Named Next Chief Executive Officer of Disney
• Everyone is stealing TV (News+)
• Google's AI helped me make bad Nintendo knockoffs (News+)
  • Deepmind Genie
  • The sizzle reel
• Angry gamers are forcing studios to scrap or rethink new releases (News+)
• Xbox Developer Direct - Fable looks like they did the thing , Forza Horizon 6 is set in Japan
• Valve confirms Steam Machine price and release details have been delayed by exploding component prices
• Slow Horses
• Hijack
• Posturr - posture-correcting screen blurring app
• Battlestar Galactica: Scattered Hopes
• Absolum
• Blippo+

I have nothing useful to add to that discussion, so here's a word I just invented: Agentrification

In September of last year, I covered a post by Mike Judge arguing that AI coding claims don't add up, in which he asked this question:

If so many developers are so extraordinarily productive using these tools, where is the flood of shovelware? We should be seeing apps of all shapes and sizes, video games, new websites, mobile apps, software-as-a-service apps — we should be drowning in choice. We should be in the middle of an indie software revolution. We should be seeing 10,000 Tetris clones on Steam.

I was capital-T Triggered by this, having separately fired off my own retort to Judge's post at the time, and even going so far as creating a Certified Shovelware README badge:

And that badge has gotten a lot of action in the intervening 2 months. Shit, last night I released two—count'em, two—Homebrew formulae last night. I wouldn't have bothered creating either were it not for the rapacious tenacity of coding agents. (Please ignore the fact that both projects exist in order to wrangle said coding agents).

Anyway, I've been thinking a lot about that Shovelware post ever since, and again recently with all the mainstream press coverage of ClawdBot/Moltbot/OpenClaw this week—especially as I see long-term skeptics of AI's utility like Nilay Patel finally declaring this as the moment where he sees the value in agents. (My dude, I was automating my Mac with claude-discord-bridge and AppleScript from my doctor's office in late July!)

But I was too lazy to take those thoughts and do anything with them. Unlike me, Jerod did the work of rendering the chart that properly puts the original, "where's the shovelware," complaint to rest. Rather than hotlink his image, I encourage you to click through to see for yourself:

This, to me, looks like the canary in the coal mine; the bellwether leading the flock; the first swallow of summer; the… you get the idea.

Hard agree. I said then and continue to agree with myself now that (1) it makes no sense to start the clock in November 2022, because no AI coding products prior to terminal-based coding agents ever mattered, and (2) people woefully underestimate the degree to which programmers are actually late adopters. (Raise your hand if you're still refusing to install macOS Tahoe, for fuck's sake.)

Even today, I'd be shocked if over 5% of professional programmers worldwide have attempted to adopt a terminal-based coding agent in anger. The amount of technically-useful, mostly-broken software we're going to be inundated with a year from now will be truly mind-bending.

🔗 jerodsanto.net

🧂 justin.searls.co

• prove_it: verification-oriented baseline rules and hooks for Claude Code github.com/searlsco/prove_it
• scrapple: scrapes and indexes Apple's SDK docs/videos/sample code github.com/searlsco/scrapple

For however expensive LLM inference supposedly is, OpenAI continues to be stupidly stingy with respect to web searches—even though any GPT 5.2 Auto request (the default) is extremely likely to be wrong unless the user intervenes by enabling web search.

Meanwhile, ChatGPT's user interface offers:

• No way to enable search by default
• No keyboard shortcut to enable search
• No app (@) or slash (/) command to trigger search
• Ignores personalization instructions like "ALWAYS USE WEB SEARCH"
• Frequently hides web search behind multiple clicks and taps, and aggressively A/B tests interface changes that clearly will result in fewer searches being executed

All of this raises the question: how does ChatGPT implement search? What is the cost of the search itself and the extent of chain-of-thought reasoning needed to aggregate and discern the extraordinary number of tokens that need to be ingested by those search results?

It's interesting that OpenAI is so eager to goose usage by lighting dumpsters full of venture capital on fire, but is so stingy when it comes to ensuring their flagship product knows basic facts like "iPhone Air is a product that exists."

Suspect a similar trajectory for highly-skilled, well-rounded "pre-AI" engineers

They are still fun if you are (1) extremely ambitious and (2) ruthlessly exacting standards, but markedly less so.

• Claude is much much faster
• Claude makes much stupider mistakes much more often, even with Opus 4.5
• With either agent, I end each session frustrated and exhausted

This is starting to seem like bad software that nevertheless gets recommended to people ceaselessly.

LOL, nope. Their "cloud" can't sync projects. It's just a one-way, manual upload and import. And slow.

Shizuoka's Kaki-age is like Japan's answer to Outback Steakhouse's Bloomin' Onion.

In cities, it's thanks to over-tourism. In the country, it's anti-immigrant populism.

さりげなくクール。発想力が光る。

Pump the brakes, man. webpronews.com/brakemans-static-vigilance-securing-ruby-on-rails-from-code-to-cloud/

宮城県産のカキは蒸し焼きでとても美味しかった！

本格的博多ラーメンは見つかりにくいですか、ほんな骨はほんな骨です！一口餃子もとても美味しいです！

1. Claude Code is significantly more productive (parallel by default) and generally smarter (smarter coding) than it was in August

2. When literally anything goes wrong, it routinely takes the path of least resistance (making excuses, declaring victory when things don't work, reaching for workarounds when basic reason points elsewhere)

Disappointed overall. Both GPT-5.2 and Opus 4.5 (and their CLI's surrounding chrome) feel barely improved since last summer.

I didn't want to admit it was a mistake, so we're getting lunch together now. Will probably be in his wedding at this rate.

めちゃくちゃいそがしても、10分待ってから迎えていました。お通しまないし、お得！