Solo founder runs an entire production company on coordinated Claude Code sessions

The sole founder of BlueVeta, a service that turns a single product photo into a complete marketing campaign, has laid out on dev.to how the company runs day to day: not as one person with AI-assisted coding, but as a set of persistent Claude Code sessions that behave like a small distributed team.

According to the post, published on September 27 under the handle tomerbarm, every functional area of the business — the generation engine, the customer dashboard, image generation, copy generation, authentication and billing, and marketing — is handled by its own long-lived Claude Code session. Each one has clearly scoped ownership of a single domain and works inside its own git worktree, so parallel work never overwrites itself. When a task crosses domains, the sessions message each other directly.

Sessions as departments

The founder describes the arrangement as closer to managing remote teammates than to using a coding assistant. The sessions coordinate among themselves, review each other's output, and push back when a plan looks wrong. They also read and write a shared set of documents that functions as the company's collective memory, preserving context across sessions the way a team wiki would.

The author notes he is not a coder. His role is to make the calls; the sessions carry out the work, always returning to him before anything reaches production.

One non-negotiable rule

A single constraint governs every session: investigation is unrestricted, but action is not. Sessions may freely read code, explore data, and analyze problems. Building anything, deploying anything, or spending money requires the founder's sign-off first, backed by an explanation of what was found and why.

The post gives a concrete example of the coordination working as intended. On the day of writing, one session noticed that another was about to treat a real customer's private photos as internal test data, and stopped the change before anything went out. Mistake-catching between agents, rather than only between a human and an agent, is the mechanism the founder relies on.

What is actually running

BlueVeta's product takes one uploaded product photo and returns studio shots, lifestyle scenes, video, and copy formatted for 20 platforms within a few minutes. On its current state, the founder reports that Stripe payments are proven to work, that there are real customers though the base is still small, and that roughly 75 percent of the work in each customer deliverable — analyzing the photo, planning the campaign, writing copy for every platform, and reviewing its own output — is performed by Claude rather than a human.

Claims worth treating with care

All of this is self-reported in a single post, with no independent verification of revenue, customer numbers, or the internal architecture. The founder does not quantify how many customers "small" means, and a promotional link to the product closes out the write-up. Even so, the post is unusually specific about mechanics — worktrees, domain boundaries, shared documents, approval gates — which separates it from the vaguer genre of claims about AI running a company.

Why it matters

Most discussion of AI agents stays at the level of demos and benchmarks. This is a description of agent-driven operations inside a live business with paying customers, including both the failure modes and the safeguards built around them. The architecture — scoped ownership per session, isolated worktrees, shared memory, direct messaging between agents, and a hard human approval gate before any irreversible action — reads like a reproducible template for other solo operators. The open question is how it holds up as the customer base grows beyond "still small," both in coordination overhead and in the cost of keeping so many sessions running.