The agent fleet

We run a small fleet of agents that do real work for the studio every day. Not demos — production work, against our own platforms, on schedules nobody babysits. There are three of them.

The three

Pax is the generalist and the orchestrator. It lives where the code lives, runs with a large set of skills, and handles the open-ended work: drafting, reviewing, planning the week, operating the CRM, answering on Telegram. When a piece of work needs judgment, it lands on Pax.

Mebo is the researcher. It runs on a dedicated Mac mini and does the slow, fan-out work — researching a story, pulling sources, drafting content that another agent will review. Keeping research off the orchestrator's plate means the expensive, latency-heavy work doesn't block everything else.

Clyde is the renderer. Also on a Mac mini, it turns approved content into finished artifacts — vertical videos with dual-voice audio, captions timed to the words. Render work is heavy and machine-bound, so it gets its own box and its own queue.

What makes it a fleet and not three scripts

Three things separate this from a folder of cron jobs.

Work is claimed race-safely. Multiple agents and multiple scheduled jobs can wake up at once. Before anyone touches a task, they claim it; a claimed task is invisible to everyone else. No two agents render the same short, no two windows send the same email.

Everything runs as a skill. Recurring jobs aren't bespoke code — they're skills an agent invokes: "fill next week's content," "research the industry," "post the weekly videos." A skill is a documented, testable unit of work, which means the same capability runs identically whether a human triggers it or a 6 a.m. scheduled job does.

Everything reports back. Work surfaces as tickets in an orchestration layer, so there's one place to see what's in flight, what's done, and what's stuck. Agents post notes; the human reads a dashboard, not twelve log files.

Why we bother

The honest reason is that it's the product. We sell systems that run unattended and only ask for a human when it matters. The most convincing way to prove that works is to run our own studio on one — the marketing, the content, the outreach, the briefs.

It also keeps us honest about the hard parts. Race conditions, idempotency, "what happens when the agent dies halfway through," cost control, the difference between a thing that demos and a thing that runs every day for a year — you only learn those by operating the system, not by reading about it.

The pieces of this fleet show up across the rest of the lab: the weekly-brief pipeline is one of its scheduled jobs, and the meeting copilot shares its auth trick.