Health apps assume you have a 22-year-old’s free time. I’m building PULS3 for everyone else.
A team of AI coaches plans your day from your Apple Health data: a workout that fits last night’s sleep, and dinner for whoever’s eating.
iPhone. 7 days free.
https://t.co/z6UOUlFbF2
@unusual_whales Canva this week: 1,271 roles cut or moved, product pulled back to Sydney. Nicely framed, but after the valuation markdowns it’s a profitability move, and AI is a big part of why: so much of what Canva does is now commoditised across the major models.
@eliadele Something we built ourselves. Each agent runs one part of the company. Best moment so far: this morning’s posting files and captions were on my phone before I’d thought to ask.
@GergelyOrosz Ask engineers what to build and you get a factory. Search is unglamorous, so it rots for years, even now that AI makes a better one a 3-hour job. Let's not even get started on X search.
The PS5 desperately needs a proper Kids Mode.
Netflix figured this out years ago. Kids get their own interface, their own library, and an experience designed for them.
@PlayStation? My five-year-old still gets an adult-oriented interface, with games he isn't allowed to play.
I started a petition asking @Sony to build a real Kids Mode, with a parent-curated library, age-appropriate discovery, and a dedicated child-friendly home screen.
If you're a parent sharing a PlayStation with your kids, I'd love your support.
https://t.co/DhodZ4Fxet
Same here. Our health coach app has a CLI that runs the real Swift core and view models, no UI. Agents replay whole conversations as test personas and grade the trace: every safety check, memory write, tool call and routing decision. The simulator is just for checking what the screen shows.
We got into the @elevenlabs Startup Grants Program.
The first place I want voice is cooking. Wet hands, phone across the bench, and you've lost your place in the recipe.
We haven't built it yet. The first test we're planning is small: the phone's built-in voice reads one step at a time, with big Next, Back and Repeat buttons. If people use that, we'll look at where a better voice helps.
What does a recipe read out loud usually get wrong?
I'm Adam. Ex-Nike platform engineer, now building PULS3: a health coach for people who don't have a 22-year-old's free time.
It's built and run by a team of AI agents: strategy, growth, building, ops. Nothing ships, posts or sends without a yes my co-founder or I typed ourselves. An agent won't act on a yes another agent passes along.
Here I'll post what breaks when agents run a real company, and the rule we add after. Like the day one agent overwrote another agent's finished plan. Now each agent writes only its own files.
@i_mika_el The coach saying "Done" when nothing had been saved. A persona caught it before any user did. The fix we're building: an action only counts once we can read it back from where it was saved; until then it shows as unknown.
AI agents can do most of the work in my company. They can't tell me what's worth doing.
PULS3 has two people. Almost everyone else on the team is an AI agent with one job: one builds, one tests, one reviews code, one measures, one coordinates. Focused like that, they're very good. One agent asked to do everything forgets, drifts and contradicts itself.
Where they fall down is creative work and strategy. A company starting from zero has no data to tell the agents what matters, so they work hard on whatever is in front of them. In one week they made about 15 changes to a feature no user had complained about. Their first two drafts of this post were about how many code changes we merged.
If you're building with agents, where have they surprised you, and where have they let you down?
Ran this against our iPhone app instead of a website: five researched personas, each on a throwaway simulator, using the real app with the live AI coach.
52 issues in a week. The one I'd never have caught myself: the coach saying "Done" when nothing had been saved.
Thanks for the prompt.
@justdesigncn I keep finding that each job needs to be well-defined, otherwise due to the inherent architecture of LLMs you get really generic output. The frontier models are very strong but you still need to tell them how to be great for your specific circumstances. And mixing contexts is out
@i_mika_el Depends on the issue. For real, showstopping issues, definitely. The QA testers have been finding issues I would consider higher priority than users are currently.
No. We barely get complaints yet, so a QA agent makes them. Five personas with sourced backstories (a rushed mum, a skeptic about pricing and privacy, a runner who tries to break things) use the real app on throwaway simulators. Their findings go in the queue, and the next round retests each fix.
@svpino The line I'd draw is between drafting and sending. Our agents draft all day, and nothing ships, posts or sends until a founder types yes in that agent's own thread. A yes passed along by another agent doesn't count.
@rauchg The win is that the reviewer isn't the author. In our codebase a separate agent reviews every pull request and reruns the full test suite on a different Mac before anything merges.
Health apps assume you have a 22-year-old's free time. I'm building PULS3 for everyone else.
A team of AI coaches plans your day from your Apple Health data: a workout that fits last night's sleep, and dinner for whoever's eating.
iPhone. 7 days free. https://t.co/z6UOUlFbF2