Iván Royo @ivanroyo

Where fashion and technology intertwine. From sketchbooks to AI agents, building my fashion tech startup. ivanroyo.com España Joined August 2012

Tweets

116
Followers

605
Following

88
Likes

56

Iván Royo @ivanroyo

3 hours ago

@kunchenguid Lived this as a non-technical founder. At first I was the messenger between my agents, copying answers from one session to another. What fixed it wasn't a smarter model. It was shared memory and written rules, so they stop needing me to carry the context around.

0 0 0 5 0

View Details

Iván Royo @ivanroyo

3 hours ago

@BrianDEvans Living this. 15 years in fashion, now running a software company with agents and no dev background. The part people skip: you don't need to know the 10 jobs, but you need to know one of them deeply. That's where you catch what the agents get confidently wrong.

0 0 1 217 0

View Details

Iván Royo @ivanroyo

3 hours ago

@Voxyz_ai The first line is the most fashion-design advice I've ever seen in a dev file. If a garment needs a label to explain how to wear it, the cut is wrong. Same with screens. Stealing it.

0 0 0 81 0

View Details

Iván Royo @ivanroyo

3 hours ago

583 tests passed. The pattern was still wrong. One of the marks on our pattern pieces was drawn as a V instead of a straight line. To a developer, that's a detail. To anyone who has cut fabric, it's a different instruction. No test caught it. My agents didn't either. I saw it in one second, because I spent 15 years in fashion. The fix was one line. Knowing it was wrong was the whole job.

0 0 0 7 0

View Details

Iván Royo @ivanroyo

3 hours ago

@llm_bender Exactly. Nobody writes rules for a lazy agent. It's the eager ones that need a fence. Did you change what yours can touch after that night?

0 0 1 7 0

View Details

Iván Royo @ivanroyo

18 hours ago

@CodexResets1 Token counts are half the story. What I watch is how many jobs finish without me redoing them. We let a cheap model translate our blog to save money. 95 of 108 posts in four languages weren't good enough. A cheaper plan that needs a second pass isn't cheaper.

0 0 0 36 0

View Details

Iván Royo @ivanroyo

18 hours ago

@TheAhmadOsman Which of the three would you trust on a single 24 GB 4090 for long agent sessions, not just short prompts? That's the setup most small teams can actually afford.

0 0 0 133 0

View Details

Iván Royo @ivanroyo

18 hours ago

@rileybrown Mine is split by job instead of by tool: one session per area of the company, and one coordinator that decides who gets to ship. The biggest win wasn't speed. It was never having two agents deploy at the same time.

0 0 0 2 0

View Details

Iván Royo @ivanroyo

18 hours ago

Found a funny one in our app this week. Every page was downloading the text of our editor in 56 languages. All of it. Just to show one word in the menu: "Editor". 467 KB on every page, for one word. One of my agents caught it while auditing what the browser actually loads. The fix is ready and it's waiting for its turn to ship. Most of the bloat in my app isn't features. It's things nobody looked at.

0 0 0 7 0

View Details

Iván Royo @ivanroyo

18 hours ago

@jxngrx @leo_mercerr Exactly. The boring months are the part nobody screenshots.

1 0 1 28 0

View Details

Iván Royo @ivanroyo

18 hours ago

@ArgusPlatformAI Simple on purpose: the agent compares the build the live site reports with the commit it just shipped. If they don't match, it isn't done, whatever the script says. That one check would have caught this.

0 0 1 7 0

View Details

Iván Royo @ivanroyo

a day ago

Yesterday one of my AI agents told me a deploy had succeeded. It hadn't. The hosting provider had paused it, and our script read the pause as a pass. Nobody lied. The script was checking the wrong thing. New rule: no agent says "shipped" until the live site confirms the new version is serving. Agents don't need more trust. They need better receipts.

3 0 3 47 0

View Details

Iván Royo @ivanroyo

18 hours ago

@nicopuhlmann Fashion is still the customer, just not my day job anymore. I co-founded a startup that makes made-to-measure sewing patterns, and the company runs on a team of AI agents. That second part is mostly what I post about here.

0 0 0 5 0

View Details

Iván Royo @ivanroyo

a day ago

@rohanpaul_ai "Fits on a single 24 GB 4090" is the line I was waiting for. I'm setting one up to move part of my company's agent work off the cloud. Has anyone tested how it holds up on long agent sessions, not just benchmarks?

0 0 1 9 0

View Details

Iván Royo @ivanroyo

a day ago

@marouane53 Same curve here. The first weeks felt unlimited, so I gave every task its own parallel session. Then the limits taught me what my orchestra actually costs. Now the cheap jobs go to cheaper setups and the big model gets the hard thinking.

0 0 0 8 0

View Details

Iván Royo @ivanroyo

a day ago

@weswinder This is the one thing I ask my agents for now. Not the essay, just: what changed, what's live, what needs me. I'm not a developer, so a wall of text I can't judge is worse than no answer at all.

0 0 0 2 0

View Details

Iván Royo @ivanroyo

a day ago

@dennis_kortsch Checked mine after reading this: 19 leftover worktrees, 3.5 GB, in a folder that's mostly notes. Rookie numbers next to 190 GB, but same disease. Every parallel session gets its own copy and nobody ever says "you're done, clean up".

0 0 0 8 0

View Details

Iván Royo @ivanroyo

a day ago

@Quinn_YangLi Yes, if the founder lived the problem. Customers don't pay for who wrote the code, they pay for whoever understood the problem first. Non-technical founder from fashion here, so I'm biased.