Fixing bugs while you sleep

The Monday meeting arrived. Two hours, blocked, no agenda title except the CTO's calendar invite which said, simply, "The Image Thing."

I had spent the weekend preparing for both possibilities. Computer vision is a real and complex discipline involving model architectures I would need to explain carefully, set realistic timelines for, and probably source a third-party API to handle, because we are not training our own vision model on a budget that also includes a Nespresso machine and whatever Marcus expensed in Bristol last October. The other possibility was that he wanted the logo bigger. I had a Figma file open just in case.

He arrived with the energy of a man who has slept very well. He did the gesture again — the circular one, at the air — and then said he'd been reading about agents.

I waited.

He'd seen something about Cursor. The new Automations thing. Coding agents that wake up on their own when something changes in the codebase — a Slack message comes in, an incident fires, a timer goes off — and they just start fixing. No prompt. No monitor. The editor, acting unilaterally, in the night, while you're asleep.

He thought this was tremendous.

I asked what this had to do with the image thing.

He looked at me the way you look at someone who has asked a question so basic it has temporarily confused you. He said he wasn't sure the image thing was the direction anymore. He said maybe the direction was agents. He did the circular gesture again. He said the word "agentic" once, carefully, like a word he'd learned phonetically.

I wrote "agentic" in my notebook. Not because I needed to. Just to have something to look at.

---

Cursor's Automations are genuinely interesting. I want to be clear about that. The idea that your development environment can respond to codebase changes, Slack messages, incident alerts — that it becomes a thing that acts rather than waits — is a real shift in what the editor is. It's not a tool you use. It's a system that runs. The distinction matters and most people making decisions about it are not spending a lot of time on the distinction.

The CTO had identified this technology and found it exciting in the way he finds most things exciting: completely, immediately, and without the subsequent steps. He had not asked what we would point it at. He had not asked what codebase events it would respond to, or what our incident pipeline even looked like, or whether Marcus would be informed before or after the agent committed something to main.

That last one I am genuinely uncertain about. The agent question, I mean. Not the Marcus question. The Marcus question has a known answer and the known answer is after, always after, and usually when the pipeline is already broken.

GPT-5.4 came out this week as well. Fewer factual errors in reasoning and coding, paired tools for spreadsheets and documents, real integrated workflow stuff rather than benchmarks nobody uses in production. I showed Priya the spreadsheet integration. She looked at it for a moment, then looked across at me, then back at the screen. She said "Marcus is going to find out about this." I said I knew. She said "today." I said probably yes.

He found out at 11:40. There was a message in the team channel in full capitals that I will not reproduce here in full but which contained the phrase SPREADSHEETS ARE SOLVED and a string of fire emojis that took up two lines on desktop.

The CTO's meeting ended without a decision. The image thing remains unresolved. The agent thing is now also unresolved. We have two unresolved things where we had one, which I am told is not how roadmaps are supposed to work, but I have looked at the roadmap and I think the word "roadmap" is doing a lot of heavy lifting in that sentence.

I have a Cursor subscription. My editor has not yet fixed anything in the night. I'm not sure if I'm relieved or if I've just not given it enough chances yet.