I Can't Tell You What Engineering Will Look Like in September
People keep asking me, “Where do you think we’ll be in six months?”
I used to try to answer. I’d hedge, qualify, offer some version of “well, if current trends continue…” But I’ve stopped. Not because I don’t care, but because I’ve realized I have no idea. And neither does anyone else.
The proof? Look backward.
6 Months Ago
In September 2025, if you’d asked me what the next half-year would look like, I would’ve gotten it embarrassingly wrong.
Claude Sonnet 4.5 and Haiku 4.5 shipped in the fall, and Claude Code hit $1B ARR. A CLI tool. A billion dollars. GPT-5 dropped in August, and then GPT-5.1 followed three months later. Gemini 3 topped benchmarks in November with Flash close behind, Google moving faster than anyone expected. GitHub Copilot launched Agent Mode and a multi-agent platform in February 2026 as a product, not a research preview. Cursor, Windsurf, Devin all went fully agentic. The “AI autocomplete” era ended without a funeral. Cognition had already acquired Windsurf back in July 2025, consolidating the agentic IDE space before most people understood it existed.
Any one of those would have been the headline of the year. They all happened in the same six months.
Today
Last week I onboarded into a new codebase. I connected Notion AI to our Slack and GitHub through MCP, asked it questions about the architecture, the team’s conventions, the open PRs. What used to take me a day of reading docs and Slack history took maybe thirty minutes.
The signal-to-noise ratio was overwhelmingly in its favor. The imperfections are real, and they matter less than I expected. That’s the thing that keeps catching me off guard: not that the tools are flawless, but that the bar for “useful” is so much lower than “correct.”
The Articles Trying to Name It
There’s a growing body of writing trying to put language around what’s happening. Boris Tane argues the SDLC is dead, that the neat sequential model of plan, build, test, deploy collapses when an agent can do all four simultaneously. The Latent Space crew is making the case that code reviews are dead, that we should be reviewing specs and intent rather than diffs. And then there are background agents, autonomous systems that pick up tasks, run overnight, and open PRs while you sleep.
I don’t think these are hype. I’d call them unfulfilled reality. The direction is visible. The tools exist in some form. But the gap between demo and daily driver is real, and nobody can tell you when it closes. Could be this quarter. Could be two years.
An Interesting Time to Be Alive
Here’s where I’m supposed to tell you what to do about all this. Learn Rust. Master prompt engineering. Build agents.
I’m not going to do that.
What I will say is that something Kent Beck said at the Pragmatic Summit has been rattling around in my head. In a joint declaration with Laura Tacho and Steve Yegge, he pushed back on the hype with a reminder that organizations are constrained by human and systems-level problems first, and that skepticism is a feature, not a bug. Gergely Orosz wrote it all up, and it’s worth reading in full.
But here’s the thing: I’ve also started being skeptical of my own skepticism. Every time I’ve thought “okay, this is where the hype outpaces reality,” reality has caught up faster than I expected. That doesn’t mean it always will. But it’s made me genuinely humble about where I draw the line.
I don’t know what engineering will look like in September. I don’t know if I’ll still be writing code the way I do today, or reviewing PRs the way I do today, or onboarding the way I did last week. And I’ve made peace with that uncertainty in a way I wouldn’t have predicted a year ago.
What I know is that this is the most interesting stretch I’ve experienced in this career. Not the most comfortable. Not the most legible. But alive with possibility in a way that’s hard to dismiss, even when you want to.