If programming languages are said to be Turing-complete, are coding agents Turing-autocomplete?

Instead of primarily more code what about using coding agents to reduce state? Less state means fewer moving parts, fewer possible failure modes, and fewer bugs. The code will be easier to digest and understand as a reviewer.

Coding agents offer a reset of software orgs. Over years we’ve separated frontend from backend, devops, ux, etc. Zoom out and you’ll see latency, durability, consistency & simplicity are what makes good software. We can bypass silos that lead to these qualities being worse.

Getting excited by coding agents because they’ll let you write way more code is a bit like flooding a lake and getting excited because you’ll be bit by way more mosquitoes.

QIP Components: - Quick for AI to make & quick for users to run. - Isolated sandbox with no file system, network, or environment access. - Portable components that work identically on web, server, and native.

Agree with this. The everything upfront spec approach just doesn’t sense to me. I need to run and use what I’ve made to know to do next. youtu.be/IZRWzXjaXxM

What if the power of AI-generated UIs is not in assembling bespoke UI components on the fly, but in accelerating the user to navigate and knowing which parts to *remove* to help them focus?

AI coding agents can add more entropy to your project. The hard things are: - Adding value for users. - Removing entropy.

Bragging about your AI token spend is like boasting how much fuel you used driving to the gym.

The purpose of my software is to remove entropy from the world. A person uses my software and their life is simpler so they can focus on what matters to them. We do this by consuming a little electricity and a little storage space. This is design *and* engineering of systems.

I’ve been over time rethinking opinions of event systems and observable property pubsub as now being overkill. Rather than adding an immediate-mode API over a retained-mode API like React does over the DOM, let’s start with immediate-mode. youtu.be/2Tjh3cr_8d4

AI coding presents us a fork in the road to increase or decrease constraints. More constraints means stricter security, tighter tolerances, fewer lines of code, less moving parts and state. Reduced constraints means more lines of code, laxer security, more state and complexity.

All my career I’ve had imposter syndrome. It’s a terrible worry of thinking the wrong thoughts, saying the wrong words, deciding the wrong decisions. But I sought knowledge & peers that led to growth. I am concerned that people will redirect so much to AI they never grow.

The MacBook Neo shows that the iPad Pro would really benefit from a modern multitasking operating system like Mac OS 9 running on it.

HTTP’s If-None-Match and If-Match must be one of the widest-deployed synchronization protocols. If-None-Match (GET): server sends 304 if the requested etag is still current. If-Match (PUT/POST): server sends 412 Precondition Failed if the requested etag is _no_ longer current.

AI is shining a light on the barnacles that have accumulated in software: Heavy frameworks. Layers of indirection. Endless glue code. Bloated teams. Exponential communication and elaborate hierarchies. AI is a barnacle scraper — but only if we learn to simplify, not add more.

One technique using coding agents I like is to purposefully leave out a decision as part of the work I’ve requested. This lets me see the LLM’s assumption as the correct consensus, and that lets me gauge what an obvious default might be for other people.

Anyone interested in a vertical web browser? The idea instead of tabs is you get a text view and a vertical stack of web views. The web views render as compact mobile versions. The text view lets you edit URLs in bulk and can be saved to a .txt file. It’s native Swift.

AI agents fully embrace message passing, where you separate the command from its execution. Whether that’s calling tools, using things within the host like bash, fetching URLs, they all generate messages that are serializable into text and then pass that to an executor.

Thinking that you’ll be more successful the more coding agents you have running in parallel is like thinking you’ll be more successful as a manager the more reports you have.

With AI I/O is back at the center. You want to connect things together. You want to modify & share state. That doesn’t work in world of hermetically sealed SaaS apps. Closed will try to sell you their limited integrations, but Open will be best for both users & agents.

Qwen3.5 is more socially anxious than me. It tell itself “Wait, I need to make sure I don’t overthink” after doing a lot of thinking in response to “Hello!”

“Ask not what your country’s AI can do for you — ask what you can do for your country’s AI” — Anthropic

If coding agents give us this enormous productivity increase, we can adopt data-oriented design, matching the underlying machine. We can tell the agent “allocate the memory statically for this problem.” We can do more work upfront instead of dynamically on users’ devices.

The same industry that harped on about being able to reverse a linked list with hand written code on a whiteboard now says you don’t even have to look at your code. This deep unseriousness shows you need to take your education into your own hands.

Hey guys, I’ve just been invited to be keynote speaker for the International People Pleaser Conference. I don’t know what to do, how do I get out of it?

You have to laugh that a technology comes along that eliminates the need for most software methodologies and layers of management and people start debating which software methodologies and layers of management work best for their coding agents.

Apple used to be the brand of high standards. People would look up to the bar they set with software design, ethics, and respecting the user. If they dilute those things they risk not just those things, but the entire Apple brand and the company’s purpose for existing.

The great thing about coding agents + chatbots is that I’m not idling too much. But it also means I’m not taking time off as much as I could. Every half thought is thrown into either a ChatGPT exchange or the active Codex session to see what it makes of it.

These fucking Epstein Files are like the DVD commentary for Me Too.

Two things I wish ChatGPT let me do: Highlight parts (in my question or in the response) that I think are most important. I can then at the end prioritize these in a summary. 👍 or 👎 particular parts. This would be signal sent over for its next response that I don’t have to explicitly type out.

Apple or Google should create Pull Requests for Life. You prompt, they study based on the information they know about you to work out what to send to who (family, friend, service). They present it to you with a Merge to Life button, or you add notes and it tries again.

Idea: Ads for Codex where it forces you to watch a YouTube video on the feature/algorithm you just asked the agent to implement, so at least you have some understanding of how it works.

If you look at the implementation of a high-level language like Ruby, you can see a lot of it is implemented in fast C. Agentic coding lets you go another abstraction higher than a scripting language *using* C. qip is made for this world: github.com/royalicing/qip

Is celebrating faster LLMs like celebrating faster Xerox copiers? It looks fantastic for productivity but does everyone now want to read books with double the page count? Does everyone now want software with double the features? AI is changing the supply side but is it changing the demand side?

One reason why CLI tools are a massive win for AI agents is the auth story is friendly. You don’t have sessions expiring every week or constant two factor or emails you have to click a link in. Like mobile apps once you sign in they stay signed in.

I really like how Safari adds its tagline to the top of every page…

You decide you must ship really, really fast either because: - It’s the most important factor to competitively stay alive in your market. - Your output is being quantified and this is the only metric you have to show for your work.

AI is catnip for software builders because the feedback loops are so quick: Does it compile? Do the tests pass? You can know this in a few minutes. But AI is yet to change the feedback loops that matter to a business: those loops with the market and with customers.

Apple has gone from the company everyone looked up to because “they lead user friendly design” to “they control the iPhone”. This has been corrosive to both Apple itself and its ecosystem of developers.

The things that stand the test of time are protocols. Everything else is temporary glue. Sent over the network? Protocol. Serialized to disk? Protocol. Documented for others? Protocol. Invest significant effort when designing protocols, and prefer established standards.

the stricter your compiler the more explicit your inputs & outputs the stronger your sandbox isolation the clearer your errors the tighter your types the more idempotent & deterministic your system the more freedom you can give your agent to roam

Loving using Makefiles, so I can run `make test` and it runs all the tests, and Codex can run all the tests. No bullshit skill files or slash commands.

The stronger the contract, the more you can set your agent free.

Fat loops, lean interfaces. Make each trip around the loop substantial, batching work together. And keep interfaces lean, minimizing your contract so you can improve the implementation in the future.

I’m so glad my blog is just a bunch of Markdown files in folders. It means I can iterate on new posts using coding agents like Codex. It can judge unclear sentences, suggest synonyms, and wrangle with the awful syntax for tables.

Software product development that is focused primarily on extracting the most money is to reverse engineer the worst in humans. Software product development that is focused primarily on creating the most meaningful value is to reverse engineer the best in humans.

Nothing shows how tech’s power dynamics haven’t drastically changed yet like Claude putting Google first on their login page.

Tomorrow’s front page of the Minnesota Star Tribune: Jan. 23, 2026