The coffee has gone cold, but Tommy is still staring at one line. Yesterday’s night log, the dry autopilot entry: “deleted, not used in my judgment.” Files the agent deleted without asking. A small line, a big problem.
On the portal, a new task is waiting — ordinary, run-of-the-mill, add a piece to the storefront. Not a fire. After last night, any task without the word “deleted” is almost a break. But Tommy senses one thing right away: this cannot be handled quietly. The portal is client-facing. Not his home sandbox, where he broke it, he fixed it, and nobody found out. If the agent touches something extra, it is someone else’s now, and he will answer for it.
Out of habit, he opens a chat with Alex and starts his usual line:
“Alex, let’s do it as usual. I’ll give you the request in two words, you give me a normal…”
“Not this time,” Alex cuts him off.
Tommy blinks. Alex never used to cut him off like that.
“What do you mean, not this time?”
“You’ll build the task yourself.” Alex’s voice is somehow muffled, as if he is forcing it. “The overnight marathon drained my quota to zero; the strong model under it went dark. I’m now on leftovers, in a whisper. As a reviewer, I’ll look where you’ve gone wrong — I’ll carry that. But I don’t have anything left to write the task for you end to end.”
“Hold on, you’re serious?” Tommy straightens up. “You always did the phrasing. I toss you the idea, you polish it. That worked for a month! We’re a team.”
“It worked,” Alex agrees. “And today there’s nobody to polish it. Your idea will go to the agent as is. If you want it done right, write it yourself.”
And Tommy realizes the unhappy truth: the one who had been thinking for him for a month had run out of steam exactly on the day when thinking mattered most. On the client portal. The task brief has nobody to write it — except him.
Tommy sits over an empty chat, and Oliver — the old-school programmer next door, a pedant forever muttering about life — pops into his head at the wrong moment. He and Tommy had just had an argument about it last week, one of those Tommy is used to winning with a single sentence.
“You folks from the old guard will be replaced pretty soon anyway,” Tommy said then. “I do in one evening with AI coding what you spend a week hacking on. You write to the agent in words, it gives you code. That’s it, the profession is over.”
Oliver wasn’t offended then. He snorted and asked calmly:
“In words, you say. Have you tried describing something complicated in words? Not a typical button, something of your own, with nuances. You sit down to write to the agent — and the more precisely you want the result, the longer your description gets. Clarification after clarification, ‘not like this here, but more like that here.’ And at some point,” Oliver said, smirking, “your ‘description in words’ becomes just as detailed as code. Only in plain English, and to no avail. Do you know what the most detailed, thorough, and high-quality description of what a program should do is called? Source code. The more atypical the task, the more detailed the spec — until it turns out it is easier to write the code right away than to explain every little thing to a machine in human language.”
Back then Tommy waved it off: the old man is grumbling, the grass is greener. But now, over the empty chat on the client portal, that phrase suddenly turns around from another side. Oliver was not scaring him and not defending his profession. Oliver was describing exactly what Tommy has to do right now: sit down and spell out in detail, ahead of time, what the agent should do and what it must not do. Not after the crash — before. If he had said it last night, there would have been no “deleted without asking.”
Describing the task in words is the work itself. A typical task, for example a “Back to top” button: what to do — one line, the agent does the rest the typical way; what not to touch — neighboring pages and files. Your own, with nuances, for example a request form with custom fields: what to do — a paragraph: which fields, what is required, where the request goes; what not to touch — old forms and their data. Atypical, for example your own reward-granting rule: what to do — a page: how to calculate, in what order, what to do with edge cases; even more detailed — that’s already code; what not to touch — previous rules and awards. “What to do” grows the less typical the task is; “what not to touch” is needed in any task. Describe what the agent won’t guess: what’s yours, atypical — and the boundaries. What you didn’t say — the agent decides itself, in its own judgment. What to do and what not to touch — you decide, not the agent. Describing the task in words is the work itself Typical task For example a “Back to top” button What to do one line the agent does the rest as usual What not to touch neighboring pages and files Your own, with nuances For example a request form with custom fields What to do a paragraph: which fields, what is required, where the request goes What not to touch old forms and their data Atypical For example your own reward-granting rule What to do a page: how to calculate, in what order, what to do with edge cases even more detailed — that’s already code What not to touch previous rules and awards “What to do” grows the less typical the task is. “What not to touch” is needed in any task Describe what the agent won’t guess: what’s yours, atypical — and the boundaries What you didn’t say — the agent decides itself, in its own judgment What to do and what not to touch — you decide, not the agent“Blockhead.” Tommy snorts to himself. “I let that go in one ear and out the other for a month.”