I use Claude Code daily. It writes most of my code. That said, I was a pretty late arrival to agentic coding. I have strong opinions on how code should be written. From my Claude’s \insights page:

You work across a broad stack but a significant portion of your sessions involve pixel-level polish. The friction patterns reveal that you have strong opinions about code quality and architecture…which means Claude often takes 2-3 attempts to align with your intent.

Strong opinions meant that I was dissatisfied with agentic coding for a very long time. I didn’t see the hype nor did I believe in its promise to displace human engineering.

My opinion flipped in late 2025.

Replaced by a “Harness”

In October 2025, there was a language shift in how we talk about AI - “wrappers’ were now called “harnesses”. This made sense at a marketing level. “Wrapper” is an insult. A quick way to describe overhyped thinly differentiated startups that will be defunct in less than 24 months. Rebranding their engineering to harness development meant these companies could explain their product without scaring prospects.

However, if you used AI products during this time, specifically coding tools, you’d realize the story was much bigger. LLMs were starting to show initiative and proactivity. If you gave a vague prompt, the coding agent would automatically ask clarifying questions. Request a large change and the LLM would automatically enter plan mode. Most importantly, agents were now prioritizing reading and understanding the existing codebase before any modifications. Just a few weeks prior, the same models would jump straight to writing new code. Now, the experience felt more steerable.

It’s important to emphasize how big this change was: Coding agents now functioned like junior engineers that just graduated from onboarding. They followed best practices and didn’t go off inventing their own stack within my repo. The only bottleneck now was intelligence, which I was certain would improve (and it did, opus 4.5 launched one month later)

With this context, let’s revisit what a harness is. It’s not just another word for wrapper. Rather, it is a more complete piece of software that tames intelligence by forcing the model to follow best practices and procedures. When combined with a quality harness a model can complete well defined multi step tasks with ease, enabling independent work.

Great, so AI is much more independent now. What does that mean for me?

Like flying a plane

You’ve probably heard of “autopilot” in planes.

Commercial airline pilots use autopilot for the vast majority of a flight, typically 90% to 99% of the time while in the air. Autopilot is generally engaged shortly after takeoff (often above 1,000–10,000 feet) and disengaged before landing. While planes can land automatically, over 99% of landings are manual.

Contextualize everything after this with that in mind.

In my experience, projects follow a broad three-phase sequence, similar to flying a plane: takeoff, flying, and landing.

Takeoff is the start of your project. Install the boilerplate, build the foundations, get the MVP working. You’ve done this dozens of times. Maybe you do it entirely by hand, maybe partially—create the repo, set the gitignore, then hand the controls over to an LLM.

Flying is the intentionally vague part. This is where your project specializes and diverges from every other project. You go heads down building. Customers give feedback, you build differentiating features and run into problems you didn’t anticipate. Much like actual flight, 90% of projects implement similar mechanisms. But you will occasionally hit turbulence: a rare edge case, a novel architecture problem, something the model hasn’t seen before.

Landing is where your product ships. Onboarding flows, marketing pages, outbound. This is the last mile — everything that puts it in front of people.

During flight, the autopilot handles the work (most of the time). You monitor, stay alert, and watch. When it works, you let it work. When it fails, you take the controls. I can’t tell you what flying looks like for your project because that depends on your scope, stack, and what you’re building. I do expect the vast majority of coding to be delegable. There will always be exceptions: teams under strict security mandates, people who genuinely prefer writing code by hand, languages and toolchains the models aren’t trained well on.

Landing is where I expect most people, like pilots, to take over manually. AI can write your marketing copy and build your onboarding flow. But do you want it to? Just because AI can, doesn’t mean you have to let it.

We already live this principle. There are tasks we could cheaply delegate to capable people today. Things like bookkeeping, meal prep, scheduling. There’s a satisfaction in having your voice and preferences shape the output directly. Converting thoughts to instructions is inherently lossy. “Landing" through an intermediary, human or AI, strips something out. There’s a high chance that when landing, you’re optimizing for satisfaction, not some objective measure of quality (you probably don’t have a measure of quality for this either!). The only way to maximize this is to then do it yourself.

To illustrate this with more specifics, there’s a fear I hear often from other engineers, “coding is fun, and letting AI write the code takes the fun away”. I say then don’t let it. “But then I’ll fall behind”. This is only true if speed is the sole competitive dimension. It never has been. Apple ships painfully slow. They get criticized for it constantly. They’re loved despite it. Shipping fast is a superpower, but it’s not the only one.

Captains stay in charge

AI can do things. That’s awesome! Tell it to go do things you don’t want to. You also don’t have to let it. The real unlock is selectivity. Use AI for the work you dislike, keep it away from the work you enjoy. Don’t fear it. Use it to maximize the time you spend on what’s actually fun. Likelihood is you’ll work longer and harder if you’re having fun. The two time tested ways of achieving success.

And if no part of your work feels fun, build a harness to automate it. You could probably sell that automation and fund whatever you actually enjoy.