Hello and welcome to the 15th issue of AI Powered Engineering, and my first issue of 2026!

I know that many of my readers, like myself, have been using Claude Code for the last 8-10 months, but something happened over the winter break - the rest of the world discovered it…and they like it, a lot. This tweet pretty much sums it up:

I also think this is also a wake up call for engineers who have been using Claude Code just for coding, who now realize - oh wait, I can use this for a whole lot more than I’m using it for now. And the nice benefit existing engineers have is that they can really cook with it. From ordering your groceries to planning travel, canceling unused services to planning your day, Claude Code is the everything machine.

The mantra going into 2026 - Claude Code is for Everyone, and yes, for those who read Lenny’s Newsletter, you know he was nice and early predicting this in a post back in October of last year.

Oh, but that’s actually kinda old news now, because there spotlight has quickly shifted to - Ralph. And if you haven’t heard of Ralph, well, then you probably should spend a bit more time in dev communities, anywhere, because it’s all people have been talking about for the last few days. Seriously, I can’t go anywhere without people raving about Ralph.

And I was pretty excited to see that my buddy Ryan had a tweet go absolutely parabolic, about Ralph. So if for some reason you’ve been missing all the buzz here, stop everything and read this article on X:

While I typically like to keep this newsletter relatively short and sweet, there’s just been too much good stuff this week…so I went a little overboard, but I think we can all feel it - 2026 is going to be a little overboard, and that’s a-okay, so let’s rock 🤘

Agentic Coding, but Unix-Pilled

Here’s a fun concept - if you want AI agents to actually be useful, design them like old-school Unix programs. Do one thing. Do it clearly. Pass clean outputs to the next step. The moment an agent becomes a sprawling “figure it all out” blob, it starts drifting, hallucinating, or confidently doing the wrong thing. This framing quietly pushes back against the idea that smarter models alone will save us.

What makes this so interesting to me is how it reframes progress. The future of agentic coding probably isn’t one mega-agent with godlike powers, it’s lots of small, dumb(ish) agents wired together with intention. Constraints aren’t a limitation here, they’re the feature. This feels like one of those ideas that sounds conservative until you realize it’s exactly how every reliable system has ever been built.

What Real Claude Code Usage Actually Looks Like

Multiple Claude sessions running in parallel is the jam now, one just isn’t enough. And as a Star Trek fan I can’t help but think of the hive mind with shared markdown files acting like institutional memory. Planning first, execution second. Constant correction. The model isn’t treated like a genius — it’s treated like a fast, occasionally reckless coworker who needs structure.

I think the key takeaway here is that good AI coding isn’t about better prompts, it’s about better workflows. The people getting outsized value aren’t asking better questions — they’re building feedback loops where mistakes show up quickly and don’t compound. This is less “AI magic” and more “AI ops,” which is probably where all of this is heading anyway.

Where the Value in Software Is Moving

There’s a quiet but important shift happening right now, and if you’ve missed it - Karri breaks it down really well in this article on X, the TL;DR is: writing code is no longer the center of the job. As models get better, code itself becomes cheaper and more abundant. What’s getting more valuable instead is deciding what to build, how to frame it, and how to translate messy human intent into something machines can actually execute.

I think way too many people are seeing this as an anti-engineering take, but it’s almost the opposite. What Karri and so many other people are saying is that the job is moving up the stack. Judgment, taste, and clarity are becoming the hard parts. The teams that win won’t be the ones who generate the most code, but the ones who give the best direction. Clarity is king.

Giving Agents a Dashboard

I was pretty exciting when I saw this tweet from Jarrod - someone finally built the obvious missing piece: visibility. A simple HUD that tells you how much context is left, what tools are running, and whether your agent is quietly approaching a cliff. It removes that low-grade anxiety of “is this still working, or did it already go off the rails?”

What’s interesting is what this implies. Once agents operate for more than a few minutes, blind trust stops working. Observability becomes non-negotiable. This feels like the early version of what monitoring tools were to servers, not glamorous, but absolutely required once things get real.

Why Agents Lose the Plot Over Time

I don’t think a day goes by where I don’t think about memory and context windows. The real failure mode of agents isn’t intelligence, I think they’re already smarter than all of us, but their memory - well, it can drive you crazy sometimes, or at least it does me! Over long sessions they forget decisions, contradict themselves, or slowly drift away from the original goal unless something actively keeps them grounded.

That “something” is where a new category starts to form: harnesses, checkpoints, summaries, tests, guardrails. Bigger models help, but continuity matters more. Intelligence without persistence turns out to be fragile.

A little Ralph bonus for you

I know a lot of people reading this right now are probably either a few days into playing around with Ralph, or the second you stop reading this, you’re going to dive in. So I thought I’d share a little bonus from the one and only Matt Peacock.

This is a super handy Ralph script for writing tests on untested features. If you’re sick of writing tests, or if you’ve never felt like you got the gist of writing good tests in the first place. That’s okay, history has moved on, and you can too.

Honestly, there’s not reason not to test the heck out of your code, errr Ralph’s code, use this script or do a quick search and find another - they’re out there, you don’t need to reinvent the wheel!

Phew, what a start to 2026! I don’t think all of my issues are going to be this long, but there was just so much to share I felt like I couldn’t hold back.

As a quick reminder, this newsletter is free, I have no sponsors, no ads, no set schedule, and no course or eBook to sell you. I’m writing this purely because I’m so damn excited about what is happening at the intersection of AI and coding that I can’t help myself.

Okay, that’s a wrap, see you next time, whenever that is!

Secret Bonus Video 👀

This section was a hit last year - so I’ll keep it going. Greg is a good buddy of mine, his videos are some of my favs, and on this one he has Ryan join him to talk about all things Ralph. Don’t sleep on this one, enjoy!