Today the agent company went live. BlackLabel now runs as an organization of roughly 87 AI agent seats with an AI CEO, supervised by one human: me. There's a charter, a roster, an orchestrator loop, four scheduled daily meetings, a live dashboard, and — because I've learned this lesson — a kill switch.
Each agent is a defined seat with a written brief: engineering seats per product, QA seats, analyst seats, down to a janitor. Dispatches run as bounded units of work with guardrails baked in. The charter consolidates every hard rule the past month taught us: zero fabrication, ship no data, verify before claiming done, one writer per resource.
That last rule exists because of what happened two days ago.
The 27 duplicate apps
On June 16 I had five AI coding sessions running in parallel, one pinned to each product app, each editing, rebuilding, and relaunching its app for over ten hours. The result: 27 duplicate app bundles sprayed across the disk, all sharing bundle identifiers, so macOS would launch the wrong or stale copy on a double-click. Cleanup was futile while the sessions ran — quarantined copies reappeared within minutes because the fleet kept rebuilding.
The fix was structural, not just janitorial. Wind down the parallel sessions, keep exactly one canonical copy of each app, put every source tree under git (they hadn't been — edits to the product apps were unrecoverable until this week), and adopt one build script that always builds from canonical source to a single destination.
Multi-agent systems don't fail because agents are dumb. They fail because nobody defined who owns what. The company charter now defines it.
Five products take shape
The product line also settled into its shape this week: five areas, each its own native Mac app. Leads (find prospects in any market, send outreach from the client's own email). Real Estate (probate and property sourcing with enrichment). Marketing (content and video from the client's own assets). Trading (signal engines — signals only). Sovereign (the full AI assistant, running on the buyer's own login).
The strategy behind all five is the same: in each category, build and own the incumbent's engine rather than renting seats to it, and sell that ownership to the customer.
What "live" means here
The org runs on scheduled infrastructure, not on me remembering to start it. Meetings fire on a timer, the orchestrator dispatches work in cycles, and everything the agents claim is supposed to trace back to a verifiable source. "Supposed to" is doing some work in that sentence — keeping dozens of autonomous agents honest turns out to be a daily discipline, and the audits that enforce it will feature heavily in the entries ahead.
One human, dozens of agent seats, five products. Now we find out what it can actually ship.