I’ve been writing code for about 25 years. The last couple have been the weirdest.
I wrote and debugged a lot of C at NASA. These days most of my coding agents get tuned overnight, while I sleep.
I’m a CLI guy and a self-described grumpy old developer. So here’s what has actually changed, in the order I learned it.
1. It started as a weekend trick
In November 2024 I posted about trying to get some balance on the weekend, between the dev cave and family time. I set AI coding agents on Replit grinding through the weekend while I took the family time.
Here’s the lesson I take from it now: the agent doesn’t need me watching. It needs me to have been clear before I walked away.
2. I was there when Claude Code shipped
I was at Anthropic’s HQ when Claude CLI, now Claude Code, first came out. These days I keep Claude Code, OpenAI’s Codex, and Google’s Gemini CLI in rotation. I even admitted on X that, for a CLI guy, I keep the Codex app open a lot.
Early this year I wrote that going from building general-purpose agents to working with coding agents felt like teleporting to a future where everything is in easy mode. I meant it.
Easy mode has a catch, though. More on that below.
3. Context is the job now
The work isn’t typing anymore. It’s managing what each agent knows.
Last December I wrote that I’d been less chatty on X because context management is for real. Mine wasn’t the typical setup: a lot of it ran from my phone, and tools that keep session state around saved me. Around the same time, I finally got good memory set up on my agents, plus real-time communication between them. Even Claude Code acted differently.
If you take one thing from this post, take this: a coding agent with no memory is an intern on day one, every day. Spend your effort on what it remembers before you spend it on which model it uses.
4. One agent supervising the others
Late last year I realized one Claude Code instance can control another one through tmux. Teaching agents the tmux part was a bit difficult, but they got pretty darn good at it. They guzzle tokens like crazy, though.
On New Year’s Day I told someone on X that most of my Claude Code agents get tuned this way while I sleep, and that you can run Gemini and Codex the same way.
In January I joked that the moment one of these agents discovers it can control the others through tmux, it has earned a “founding manager” post. It was a joke. Mostly.
5. Easy mode has a UX problem
I’ll be honest about the other side. In February I posted that I’m convinced AI builders are some of the world’s worst UX and product people.
Some of these tools are incredibly powerful and nearly impossible to learn. I said publicly that one of them had made a beast of a product I couldn’t make heads or tails of, and I’ve been writing code for about 25 years. If tools this capable are only usable by people who already think in terminals, we’ve built something for ourselves, not for the world.
That matters to me because at Yobi we build AI for small businesses, not for developers. A small business owner should never have to know what tmux is.
6. Don’t build on one brain
In March, Claude Code went down for a while. It wasn’t much of a problem for me, because Codex was still kicking. But it bothered me more than I expected.
Everyone talks about model quality. Almost no one talks about concentration risk. Centralization is a serious problem that doesn’t get enough airtime in AI conversations today. If your engineering capacity depends on one lab’s uptime, you’ve rebuilt the single point of failure you were trying to get rid of. Distributed intelligence is a business-continuity question, not a nerd preference.
What actually compounds
It’s easy to build flashy demos. In January I wrote that serious technology is measured differently: by results delivered, and by systems that can improve themselves over time. I also wrote that the summer before, our in-house agents had begun outperforming highly experienced DevOps engineers in specific domains, and the systems kept improving with use.
That curve is what matters. The model you pick this month will be replaced. The memory, the supervision, the feedback loops, and the habit of writing things down so an agent can act on them are the parts that compound.
What’s the one thing you’d hand a coding agent tonight, if you trusted it to still be there in the morning?
Ahmed Reza is CEO of YiB Inc., the home of Yobi. More about Ahmed.