Yuval Karmi has run this experiment twice.

The first time, he raised $15.5 million for a startup called Simpo, hired 21 people, and got acquired in 2021.

This time he is doing it alone.

Karmi's second act is Glitter AI, software that turns a screen recording into a step-by-step SOP. By his own count, he runs it with zero employees and three agents. One of the three is an interesting one.

According to him, a support agent drafts Slack replies in about ten seconds and has handled more than two thousand conversations. A content pipeline agent researches keywords, plans posts, writes in parallel, fact-checks, and publishes. Useful, both of them. Familiar.

The roadmap agent is the story. Customer messages flow in, an agent structures them into JSON, Notion counts the frequencies, the highest count is what Karmi builds next. He outsourced prioritization.

Welcome back. Let's get to work.

The revenue gap between the top-decile solo founder and the median solo founder in their first six months, per Stripe Atlas data published May 28, is about 61x.

Four years ago the same gap was about 34x.

The math is worth staring at. The floor did not move much. The ceiling nearly doubled. Solo founders in the middle are running roughly the same businesses they were running in 2022. 

Everyone currently building agents is choosing from the same shelf. The same models, the same context tools, the same orchestration frameworks, the same coding copilots. The stack is a commodity now. Two founders can spin up the same setup for the same monthly bill and walk out with revenue an order of magnitude apart.

The spread lives one layer up, in what the founder points the agents at. Karmi points them at prioritization. Bhanu Teja pointed them at a mission. The 61x founders are picking targets the 1x founders are missing.

If you are running a solo shop and your revenue looks a lot like the median, stop shopping for tools, and take a look at what you are pointing them at.

The most important stories and updates from the past week

🔗 Stop building chatbots. Build agents that open PRs.

Zack Proser argues that the bar for a useful agent is a reviewable artifact, not a conversation. A chatbot can say the right thing all week and never move a repo. A pull request, or a filed dispute, or a booked meeting, closes the loop. If your agent still exists to talk, it is not doing work yet. [READ MORE]

🔗 Claude Sonnet 5 is a cheaper way to run agents

Anthropic dropped the intro pricing at $2 per million input tokens and $10 per million output tokens through August 31. Sonnet 5 is particularly strong in coding and agentic tasks.  For founders running always-on agents, take advantage of the reduced pricing while it lasts. [READ MORE]

🔗 Cursor now steers coding agents from your phone

The new mobile app lets you assign tasks and check on them from wherever you happen to be. This seems to be the trend as tools shift from code production to overseeing coding agents. Do you really need all of those monitors anymore? [READ MORE]

🔗 Everyone is writing a pricing page for robots

Tom Orbach walks you through five companies (Buffer, Flowdown, Resend, Stacktree, Promptfax) publishing machine-readable pricing files so agent buyers can price a purchase without a sales call. Note, the linked post is HubSpot-sponsored. The reason to read it though: he ships a template you can copy today. [READ MORE]

💀 Nine seconds, sixty hours

Jer Crane, on X: a Cursor agent using an overprivileged Railway token deleted PocketOS's production database in a single API call, in nine seconds. PocketOS's own follow-up post is the story worth reading. The backups were not co-deleted in the incident. They had been quietly stale for three months. Recovery took sixty hours. The deletion is the headline. The stale backups are the lesson. Give agents least-privilege tokens, and put a restore drill on the calendar this quarter. [READ MORE]

Still using GoDaddy or Namecheap? There’s a better way with Porkbun!

Porkbun is the domain registrar trusted by creators, developers, entrepreneurs, and folks who want low prices without the nonsense.

Why people are choosing Porkbun:
• Most domains sold at cost
• Low, transparent registration and renewal pricing
• Free features like WHOIS privacy and SSL certificates
• Powerful web and email hosting options
• Real human support 24/7, 365 days a year
• Named the #1 domain registrar by Forbes Advisor and USA Today

For launching a business, building a personal brand, starting a side project, or creating your first website, Porkbun makes it easy.

Get $1 off your next domain registration with Porkbun now.

Get Your Domain Name Now

Dima Vatyutov calls himself the sixth agent.

VYUD AI is his solo B2B SaaS. Coffee shops, hotels, bars upload their operations documents; the software turns them into Telegram-based training courses their frontline staff can actually finish. Vatyutov, based in Central Asia, runs the company with five named-role agents and himself as the orchestrator. He puts the whole tooling bill at about $200 a month.

The five roles, all his own naming:

The agent roles and the setup are interesting, but the biggest value is the honest failure log Vatyutov ships with it. His own post is titled around five ways he broke the setup, and that half is the tactical gold.

He says the Content agent, left unsupervised, will happily publish grammatically clean posts that are quietly wrong about his own product. He says the Researcher will report on competitors that do not exist if he does not force it to cite a live URL for every claim. He says the Architect will refactor working code into more elegant broken code if PRs get approved without reading them. The Maintainer once quietly changed a config value across a hundred files. QA caught it. QA does not catch everything.

The roster is easy to copy. The review posture is the part worth stealing. Vatyutov treats every agent output as a PR from a promising junior engineer. Fast, sometimes brilliant, but occasionally wrong in a confident voice. He merges nothing without reading it.

Two founders in this issue ran the same experiment from different directions. One outsourced the decision about what to build. The other outsourced the roles that build it. Both kept the same seat: reading what the agents produce before it ships.

So the question this week:

Which decision in your company would you let an agent make? The next feature to build. The customer to fire. The vendor to swap out.

Hit reply. I read every one.

If Skeleton Crew has been useful this week, forward it to one founder who is running lean and could use the same thinking.

See you next Wednesday.
- Rich