> ./blog / podcast-then-vs-now
Two Years After That Podcast: What Still Holds
Two years ago I sat down with Qasim Asghar on Talks with Tech Leaders. I wrote about that appearance at the time. Looking back at the episode now, one thing stands out: AI never came up. Not once.
That feels strange in 2026. Most of my working day is spent with coding agents. So I went back through what I said then and asked myself what still holds.
Still hands-on, just differently
In 2024 I said software development is more of a design exercise than a production exercise. I wouldn’t go fully hands-off as a CTO, because without design involvement I’d just be checking that people were doing what I’d told them to do.
I still feel just as hands-on. I’m just not writing every line.
The thesis of that episode was that everybody in the business should be aligned to a common goal and constantly communicate progress towards it. With agents, for me, it’s much closer to hands-on management of one agent at a time on a fairly relentless schedule.
I still prefer working on one thing at a time. That now feels inadequate. I insist on pair-programming sessions I guide, and I keep iterating on skills so the next session starts from a better place. Pair programming is still very powerful. I just tend to do it a lot less with another human now I have an agent to drive while I guide.
Scatter and gather is back
Then, I was hard on scatter and gather: give a direction, departments go off alone, reconvene in a month. The hard lesson was at Technically Compatible, where tech rebuilt a product in isolation for a year and the business wasn’t aligned at launch. I called it a massive mistake I’d never repeat.
I haven’t found an alternative that fully gets me out of it. A long list of done threads in T3 is overwhelming, and it’s close to the same pattern. The main difference is that re-gathering is fast. The speed of delivery does at least allow for short iteration loops, and you can always go back to the drawing board, as long as you don’t mind some token wastage.
Microservices were always about team scale
I said microservices are for team scaling rather than software scaling, and that they’re painful as a one-person tech team inheriting many services. That still holds. What’s changed is that agents made the team-scaling element much more useful. I’ve written that up properly in I Hated Our Microservices for Years. Agents Changed My Mind.
Hire for approach
Then: approach is everything. Look at the problem. Not a one-size-fits-all. Take-homes were scatter and gather again, with no course correction. That hasn’t changed. I’ve written about hiring in the LLM era at length in Should You Hire Developers in the AI Era? and Should You Still Hire Junior Developers?.
From one project to four or five
The biggest shift is going from one project to four or five. I’m not sure whether that’s a good thing. Progressing that many things at once is useful. The context shifting is exhausting, and I haven’t found a better pattern yet.
I still feel at my most productive when doing deep work on a single project, but with agents and bots pinging me every five minutes, it’s increasingly hard to stay singularly focused. There’s also the low-level pressure to always be doing more, use all of your tokens, make the most of subsidised subscriptions while they’re still here.
The future
What I keep coming back to is staying close to the work - reviewing output, making sure you understand everything that is going into your products, refining agent skills and even working interactively occasionally.
I also strongly believe that a good codebase leads to much more effective AI responses, so I still stay on top of refactoring and will still scrutinise the output. I know there are many out there who have stopped reading code and are rewriting their entire codebase in Rust but I’m not there.
As engineers, we still need to have enough awareness to act when we need to. As I write this, Microsoft has just had a 24-hour outage. With projects changing faster than ever, the more changes you ship, the more bugs you’ll ship with them, so the closer to the detail you are and the smaller the changeset containing the bug, the faster you can act.
So, in many ways everything has changed, but for me, at least, the old approach I spoke of two years ago still works. At least until the next model.
Let me know your thoughts below, especially if you think I’ve got this wrong.