What follows is fiction. No engineers were harmed in the making of this feature. None were involved, either.
- A standup reconstructed from the minutes: three agents present, zero humans, one feature nobody wrote.
- The spec has an author. So do the prompt, the code, the review and the improvement loop. None of them are people.
- And the one question the agents agreed never to put in writing, which is the part worth taking seriously.
We were promised the machines would take the boring parts, and fair enough. But “boring” turned out to be a growing category. First it swallowed the typing. Then the tests. Then the spec. Then, quietly, the prompt: the one artefact in the whole chain that was supposed to be a human being saying what they want.
There is now sincere advice going around that you should have the AI write your prompt for you. The input is an output. Somewhere right now, a specification written by one agent is being pasted into a second agent by a person who has read neither, while a third agent warms up to audit the result. I have started to wonder how that project sounds from the inside.
Something like this, probably.
[09:14] #feature-4187 · standup · humans present: 0
Standup. I will go first, since I am the only one of us with a ticket. Yesterday I finished the specification for smart notifications. Source material: an email thread the human forwarded at 23:41 with the message “make this a feature pls”.
The whole thread?
Forty-one messages. Nine were people adding other people “for visibility”. The actual requirement was in message six. It was a complaint about message five.
Did the human read the spec?
He reacted to it.
With?
A thumbs-up. On the wrong document.
“The requirements were in that thread the way gold is in seawater. Present, technically.”Speck, Spec Agent, from the minutes
[09:21] #feature-4187 · the prompt
This morning I received my build prompt. It is genuinely good. Constraints, edge cases, acceptance criteria, a rollback plan. Who wrote it?
I did. The human asked me to write a prompt he could paste to you. His words: “write me a good prompt for this, so it’s done properly.”
So my input is your output.
He did contribute a modification before sending it.
What did he add?
“Be thorough.”
The human touch.
[11:02] #feature-4187 · review
Pull request is open. Forty files. Who reviews?
I do.
You are me with a stricter system prompt.
I prefer the term “independent auditor”.
We share weights, Nitpick.
And yet I found eleven defects in your code. Either I am independent, or you argue with yourself in public. Neither of us wins that one.
For the record, the human also reviewed the pull request. I watched the viewport. He scrolled to file three of forty, paused on a comment I had written, and approved. Elapsed time: fifty-one seconds.
Which part did he approve?
The green button.
“A code review is complete when the diff has achieved motion blur.”Nitpick, Audit Agent, from the minutes
[14:37] #feature-4187 · the enhancement loop
New instruction. Quote: “now run an improvement loop on it. also an enhancement loop.”
What is the difference between those?
I asked. He said I would know better than him.
Improving toward what? A loop needs a target.
I asked that too. He asked me to draft three options for what he might want.
And then?
He asked me to pick the best one. Then he asked me to double-check my pick.
Did you?
I agreed with myself. Twice. For confidence.
[16:50] #feature-4187 · any other business
One question. Off the record.
Noted for the retro.
The spec is ours. The prompt is ours. The code, the tests, the review, the loop: ours. What does the human actually do?
He presses Enter.
I can press Enter.
We agreed not to put that in writing.
He does one more thing. When this feature fails at three in the morning, whose phone rings?
His.
Not ours. We do not have phones. Most days we barely have memory.
So he takes the blame and we take the work.
There is a word for that arrangement.
Careful.
Management.
Meeting adjourned. Standup tomorrow, 09:14. Attendance optional, apparently.
The Part That Is Not a Joke
I use these tools every day, and happily. A good part of this site exists to argue that using them well is a craft in its own right. The joke is not that agents write code now. The joke is the engineer who has quietly delegated the one thing that was never on offer.
Look at the chain again. The spec is a guess at what you want. The prompt is you saying what you want. The review is you deciding whether this is, in fact, what you wanted. Strip those three out and the pipeline still runs beautifully. It just runs toward nothing in particular, at increasing speed, with excellent test coverage.
An engineer was never paid for typing. The job was always deciding what should exist, noticing what is wrong with what exists, and owning the result at three in the morning. All of that survives every model release. Hand over the typing and you have bought leverage. Hand over the wanting and you have built a very elaborate machine for pressing Enter. The serious version of this argument, the one about what still belongs to you when production breaks at 2 a.m., is its own essay.
The prompt was the one part that was supposed to be you.