The Company That Lets You Rewrite the Ending Is Quietly Eating the Streaming Industry

November in Dublin, rain hitting the window facing the River Liffey. Liam O'Sullivan drank his third coffee down to the dregs, the crack along the rim bleeding another brown stain into the steam. On screen sat a pull request freshly generated by Fable 5 — a state machine for an insulin pump communication protocol. Four nested callbacks in the diff, variables named data1, data2, temp. It ran. It was ugly enough to make him clench his molars.

He hit reject, sent a Slack message to Saoirse: Don't merge tonight. I want another look.

Saoirse Flynn was the security architect at Aegis Health, seated two desks to his left. She didn't reply to the message. A few seconds later, a packet of Gaviscon landed beside his keyboard. Liam didn't look up.

"It's not my stomach."

"Then it's the code."

He didn't deny it.

Earlier that year the FDA had sent a warning letter over their implantable glucose monitor's communication protocol. The fault wasn't a logic error — it was a race condition. Under extremely low probability, two concurrent instructions would overwrite each other's checksums. Nobody had caught it during code review because it was buried inside 3,400 lines of auto-generated glue code. Everybody had assumed the AI-generated serialisation layer couldn't get things wrong.

That night Liam dragged a whiteboard into his garage and mapped the entire call chain. When he finished, he wrote at the bottom: I don't trust it. The question is how do I distrust it in a way I can trust myself.

Saoirse came looking for him at 1 a.m. She saw the whiteboard, took the marker, circled the words "don't trust it," and wrote a single phrase beside it: short leash.

That phrase became the core metaphor for their internal AI usage policy, though it would be five more months before anybody wrote it into a document.

Liam's approach had nothing to do with the YouTube videos where people set a dozen agents running in parallel. He used no orchestrator, no batch-processing of tasks, no "YOLO mode" that skipped permission confirmations.

Crush — the fork of an AI coding agent that he maintained — had one hard-coded rule: any time it was about to alter code, it had to display a diff and wait for human authorisation. He named this rule short_leash and annotated it with an Irish proverb: Trust the man, tie the horse.

Every time he sat staring at a diff, Liam read it like a stranger's commit. Each line the agent meant to change, he walked through in his head: why was the loop written like that, was the boundary condition missed, would he have chosen that variable name or had it been thrown together thoughtlessly. If he sensed it veering off course, he rejected, no explanation, just one word: again.

Once it changed >= to > inside a TCP reassembly block. One byte's difference. If he hadn't read the diff, the bug would have surfaced in the seventh hour of stress testing, right when the packet sequence counter rolled over. He stared at the single-character difference for a long time, then pressed reject, took his hands off the keyboard and drew a short leash in the margin of his notebook. On the collar of the dog at the end of it he wrote Fable.

He wrote his own task skills as well. Large tasks got broken into small steps, each step required to land in a separate commit. Commit immediately after every sub-task — not out of compulsion, but because he had once watched Opus abruptly delete a previously written authentication module on the fourteenth conversation turn, a module entirely unrelated to the current task. From then on his rule was absolute: commit after every sub-task, so that the next time it went mad it could destroy at most the current half-step, not an entire day's work.

Saoirse looked through his commit log once. One pull request broken into seventeen commits, every message identically formatted: task number, what was done, model version, whether it had been reviewed. At the bottom of that log page she wrote a single comment: You should write this into a policy.

So he did.

In July, Aegis Health published the first draft of its internal AI usage policy. Liam wrote it; Saoirse was the first reviewer. Beside the line that read "Any AI-assisted PR must be reviewed line-by-line by its human author," she wrote a single word in red ink: obviously.

Liam stared at that word for a long while.

In six years of working together, he had come to know Saoirse's sentence structures intimately. She never wrote "I think," never wrote "It is worth noting," never wrote, "It merits deep reflection." Her code reviews came in exactly two forms: either a boundary condition had been missed, in which case she would cite the exact line number and provide a test case; or the approach was wrong, in which case she would provide a counter-example with a link to the documentation. This obviously was a third thing. It didn't mean "needless to say." It meant "The fact that you had to write this down means someone out there isn't doing it — and to you and me it should be as natural as breathing."

He later added another clause to the policy: every PR description must disclose, under an "AI Disclosure" heading, the exact model name and version used to generate it. Not for attribution. So that the reviewer would know the calibre of suggestion behind the code in front of them. Code written by Fable 5, code written by Opus, code written by a human — each deserved to be read against a different standard.

Saoirse approved. The CTO approved. At the bottom of the document, Liam added a footnote: It's like walking a dog that's stronger than you. The length of the leash decides who's in charge. If the leash is longer than your arm, you are no longer walking the dog. You are being pulled along behind it.

In September, Liam finished refactoring a session-layer retransmission protocol. The core state machine was entirely handwritten; the test cases and serialisation glue had been generated by Crush in short-leash mode, every diff confirmed. In the PR description he wrote it plainly: Fable 5, Crush fork, all AI-generated portions reviewed line-by-line by the author.

Saoirse's review changed exactly one line. Inside a ring buffer's cursor-wrap condition, she changed offset >= buffer_size to offset > buffer_size. Boundary condition. She wrote no comment, only marked it approved.

Liam stared at the green tick, leaned back in his chair. At the edge of his keyboard membrane, a corner had peeled up; he smoothed it with his thumb. It peeled up again.

Outside the rain had stopped. Light lay broken on the surface of the Liffey, like a metaphor he didn't need to write.

He picked up his phone and messaged Saoirse: Coffee's on me tomorrow morning.

She replied: You know I don't drink coffee.

He replied: I know.

At 7:45 the next morning Liam set a hot chocolate on Saoirse's desk. Stuck to the lid was a yellow Post-it note with three words: Short leash.

She lifted it, took a sip, turned the note over, and drew a leash on the back. At the end of it there was no dog. Only a person.

Then she opened JIRA and began the day's first code review. Liam put his headphones on. A new diff confirmation box appeared on his screen. He glanced at it, pressed approve, and in the margin of his notebook added one more stroke beside the short leash — this time the leash held by a human hand, drawn clear, every knuckle and fingernail exactly where it ought to be.