An AI agent at Veronata answered 47 support emails yesterday. It wrote them in five languages. It refunded six users. It escalated one to a human. The whole thing cost us $1.83.

This is what "agentic" actually means. Not a chatbot. Not a copilot. Not a science demo. An agent is a piece of software that reads, decides, acts, and either finishes the job or hands it off — without a person in the loop.

Most writing about agents focuses on what they could do, eventually. We'd rather show you what ours did today.

Here are six things our agents do across the seven products in our portfolio:

  1. They answer the mail. Across StarSinger, Tuck, SnoreCam, PinkyBloom, PinkyBond, Quit336, and Tadula, our support agent handles roughly 80% of inbound email without a human reading it first. The rest gets a tag and routes to one of us. The agent has its own evaluation harness; we sample 10 conversations a week and grade them.
  1. They write App Store reviewer notes. Apple requires a one-paragraph explainer when you submit any in-app-purchase change. Our agent writes those. We've shipped 38 IAP changes this year. The agent wrote 38 reviewer notes. A human approved 38 of 38.
  1. They pick the next product. Once a quarter, an agent surveys the iOS App Store top charts, scores 200+ category-leaders for "build feasibility × portfolio fit," and proposes the next launch. Tuck was one of those proposals. So was a product we did not build because the agent's own scorecard rated it badly enough.
  1. They write the screenshots. Each app on the App Store needs localized screenshots and metadata for 30+ App Store locales. Our agent generates them, A/B tests two versions per locale, and picks the winner after 14 days using App Store Connect's analytics API.
  1. They handle the deploys. When PinkyBloom's iOS build needs to ship a hotfix, the agent runs fastlane, writes the release notes, submits to Apple, replies to the reviewer's questions, and posts a Slack message when it's live. Median end-to-end time from "merge to main" to "live in production" is 2 hours 11 minutes. The agent is asleep for most of it.
  1. They watch the books. Stripe webhook fires. Refund-threshold trips. The agent files a chargeback dispute, attaches the user's session log, and tracks the outcome. No one in the studio touches a Stripe dashboard most days.

None of this requires a breakthrough. All of it requires plumbing. The hard parts of agentic operations are unglamorous: prompt versioning, environment isolation, secrets rotation, evaluation harnesses, escalation protocols, and the boring discipline of writing down what "done" means before you start a task.

We are not telling you any of this because it is exciting. We are telling you because everyone else writing about agents writes about the future. The present is more interesting.

The present is: a small team — one human, today — operates seven consumer businesses that ship code, answer customers, file App Store paperwork, and collect money. The agents do most of it. The human chooses what gets built and reviews what comes out. The economics of this are different from a small SaaS team. We will write about those in a separate essay.

For now, the only claim we would make is this: an AI agent is not a research demo. It is a 47-emails-a-day, 38-reviewer-notes-a-quarter, $1.83-a-day kind of machine. It is the most ordinary thing we use.

We will publish numbers on this every month.