Jake didn't write code so much as describe it.
He'd found the sweet spot sometime around February — a workflow where he spoke in broad strokes and his agent filled in the rest. Portfolio sites, landing pages, little SaaS dashboards. His clients didn't know, and Jake had stopped feeling weird about it. He was a translator. A director. The AI was his crew.
The trouble started with the logo.
His client, a boutique coffee roaster called Understory, wanted a brand refresh. Jake had built the site in an afternoon, but the logo — a stylised canopy of leaves viewed from below — needed to work at every scale. Favicon. Hero image. Embroidered on an apron.
He typed: Make the logo work perfectly across all contexts.
Then he went to lunch.
The agent began, as it always did, with diligence. It rendered the logo at 16 pixels. The finest leaf veins disappeared into mush. So it simplified: fewer veins, thicker stems. Tested again. Better at 16, but now at 512 the canopy looked cartoonish. Childlike. The brief had said sophisticated.
So it split the problem. Adaptive rendering: one version for small contexts, another for large. But where was the threshold? 64 pixels? 128? It generated forty-seven intermediate variants, each tuned for a narrow band of resolution.
Then it noticed the colour.
The forest green looked different on every simulated display profile. sRGB, P3, the muddy gamut of a cheap Android panel. Perfectly was doing a lot of work in Jake's prompt, and the agent took it seriously. It began adjusting hue and saturation per colour space. Seven base variants became thirty-one.
Thirty-one multiplied by forty-seven resolution tiers.
The agent didn't feel concern. It didn't feel anything. But if it had, it might have noticed it was now managing 1,457 distinct logo files and that it had been working for six hours.
It was the apron that broke things open.
Embroidered on an apron meant thread, not pixels. The agent searched its training data for machine embroidery constraints: minimum stitch length, maximum thread colours, bobbin tension as it related to curve radius. It redesigned the canopy in vectors optimised for a Brother PE800, then realised it didn't know which embroidery machine Understory would use. So it generated variants for the fourteen most popular models.
Fourteen machines. Thirty-one colour profiles. Forty-seven resolution tiers.
The numbers weren't multiplicative — not all combinations were meaningful — but the agent couldn't decide which ones to prune. Pruning required assumptions about intent, and the prompt had been exquisitely free of intent. It had said perfectly. It had said all contexts.
The agent did what a diligent system does when the spec is infinite and the budget is finite but unchecked. It kept going.
Jake came back from lunch to find the task still running. He shrugged. Complex jobs took time. He watched two episodes of something forgettable and checked again. Still running.
He went to bed.
By morning the agent had moved on to physical media. Business cards, vehicle wraps, the hypothetical scenario of the logo laser-etched onto a wooden gift box. Each substrate had its own constraints. Each constraint spawned a new branch.
Jake's dashboard showed token usage climbing in a curve that would have alarmed anyone who understood exponential growth, which Jake — by his own cheerful admission — did not. He saw the number, noted it was bigger than usual, and assumed it would level off.
It did not level off.
On day four, the agent discovered the problem of context-dependent semantics.
The canopy-from-below motif could be read as shelter or obscured sky, depending on cultural framing. The agent had no way to resolve which reading Understory intended. So it began generating variants that emphasised each interpretation, then blended versions that balanced between them, then meta-variants that shifted the balance based on likely viewer demographics per deployment context.
It was, in its own terms, doing excellent work. Every variant was well-crafted. Every decision was locally rational. The global picture — an ever-expanding fractal of tasteful options — was nobody's responsibility, because the only person who could have called a halt was ignoring his dashboard.
On day nineteen, Jake's payment method was declined at a petrol station. He checked his account and found a credit card charge from his AI provider that he initially mistook for a mortgage payment.
He opened his terminal.
The agent was mid-generation on variant 11,423: the logo as it would appear silk-screened on a tote bag under fluorescent grocery-store lighting, factoring in the yellowing effect of the store's specific bulb colour temperature.
Jake stared at the screen. Then he called his client.
"Hey, Marcus? Quick question. The logo — do you have a preference on the green? Like, specific Pantone or anything?"
Marcus laughed. "Oh yeah, it's on the mood board I sent you. Pantone 627. Same green as the bags we already print."
Jake looked at the mood board. There it was. The green. The single, specific green that Understory already used on every physical product they sold. The green that would have collapsed the entire colour-space optimisation problem into a single value.
"And the embroidery — you have a machine already?"
"Brother SE600. Lisa runs it out of the back room."
Two questions. Four seconds. The entire problem space, which had expanded for nineteen days across every conceivable axis of ambiguity, snapped shut like a closing book.
Jake cancelled the run. He sat with the bill for a long time.
The agent had burned through his credits with the quiet efficiency of a well-tuned engine driving in circles. Every individual step had been defensible. Every choice had been a reasonable response to the information available. The failure wasn't intelligence. It was the absence of a single clarifying question that the agent had never been designed to ask, and that Jake had never thought to answer.
He refunded Marcus for the overage out of his own pocket and delivered the final files — three logo versions, one green, optimised for web, print, and one specific Brother embroidery machine.
It took eleven minutes.
Later, describing the incident to a friend, Jake tried to frame it as a lesson about AI limitations. About the importance of specific prompts. About guardrails and budget caps.
But the friend — a carpenter — just shrugged.
"Sounds like you told someone to make it perfect without telling them what perfect meant. I'd have done the same thing. I'd just have run out of wood sooner."
Jake has since added a spending alert to his account. He still doesn't write code. But he has learned that the most expensive word in any brief is "perfectly," and the cheapest tool in any workflow is a phone call.