This blog went quiet for a while, and somewhere in that stretch I moved to a bigger company.
Most of what I wrote before that was a bet, even if I didn’t call it one at the time. I bet on sub-agents, on hooks, on CLAUDE.md, and on letting non-engineers open PRs, and I wrote about each of them like I’d figured something out.
Enough has changed that I can score them now. One held up, two I’ve mostly dropped, and one turned out to be a bet about something else. Looking at them side by side, the one that held up was a way of working, while the two I dropped were files and scripts I had to keep maintaining.
Sub-agents held up
This is the one I’d make again. I still use sub-agents to keep my main context clean, to review code, and to run independent work in parallel. When I wrote about where to split work, I said the real reason they’re useful is context isolation, and that part still holds.
What changed is that I don’t build many of them anymore. Back then I was writing my own and got pretty carried away with it, whereas now most of the ones I use ship with the tool.
Hooks didn’t carry over to a second tool
I wrote two posts about hooks. I had scripts that injected git state when a session started and fed in recent errors, plus another one that rewrote risky commands before they ran.
I dropped all of it for a pretty boring reason, which is that I use Claude and OpenAI models now, sometimes at the same time. My hooks were config for one tool, and the other tool doesn’t read them. That’s lock-in on a small scale, because anything I build into one tool I either have to rebuild in the other or it turns into a reason not to switch. I’d rather have my workflow behave the same in whatever I’ve got open.
CLAUDE.md matters less than I said
I called it the foundation of every session, and I rewrote mine about a dozen times trying to get it right. These days I have one system prompt whose whole job is to cut the slop, and that’s pretty much it.
Part of it is that the model can reread the context every time, so it doesn’t need a briefing on how the project is laid out when it can just go look. The other part is that memory tends to steer things the wrong way. Anything I tell a model to remember is a bias, and the model applies it whether or not it fits what I’m doing right now. Honestly, it works like a cache that saves time until it goes stale, and I’d rather spend the tokens rereading than debug a stale instruction.
The non-engineer bet was about company size
The post about non-engineers opening PRs was all guardrails. We put CODEOWNERS on the sensitive paths, set up safe zones for low-risk changes, and opened agent PRs as drafts, all because I was worried about people who don’t know what they don’t know.
Where I am now, everyone opens PRs and people embrace it. The CODEOWNERS setup looks different and there’s more trust than there used to be, and review seems to be enough.
I don’t think my old concerns were wrong, and they’re probably still valid for a small team. But I wrote them like they applied to everyone, which they don’t. So a lot of what I called a lesson about AI was a lesson about the company I was at.
What I’d bet on next
My next bet is agents working unattended, where I watch fewer steps and check more finished work. It follows from what held up. Sub-agents worked because the task was isolated and I could tell when the result was garbage, and everyone opening PRs works because review catches what it needs to. In both cases the work runs first and someone checks it after.
I don’t know yet how much to let them do without me, so that’s the next thing I have to figure out.