caelith labs

teardown - the agent crew behind two media brands

How I run two media brands with an agent crew.

Every day a short film for @vektor.fm and, on its own schedule, one for @huberttheinventor. Agents do the repeat work. I sign off at a few fixed points. This page shows the parts, including what still fails.

01The crew: one role per agent

Each agent does one job and hands a file to the next. None of them can publish.

1 researchresearcherFinds the topic and proposes it on a card.
2 scriptscriptwriterWrites the narration. It has to sound like speech, not like a report.
3 imageart directorPlans and builds the pictures for each beat.
4 voicevoiceProduces voice takes. I pick the take.
5 cutrendererAssembles the film from script, pictures and voice.
6 releasepublisherPrepares caption and files. I upload and post myself.

02The five points where I decide

Nothing moves on until I approve. Each approval is tied to the exact file I looked at: if the file changes afterwards, the approval no longer counts and I have to look again.

  1. Proposal. Is this the right topic for today?
  2. Script. Does it say something true, and does it sound like a person?
  3. Silent cut. The film with pictures and no voice. Does it work without sound?
  4. Voice take. I pick one of several takes.
  5. Render and cover. The finished film and its thumbnail.

Confirmed by the operator 13.10.2026 for @vektor.fm. Upload and posting stay manual.

03The reviewers

Separate reviewer agents score the work before it reaches me: one for the script, one for the opening seconds, one for retention, one for the visual rules. A film that fails a review goes back a step instead of forward.

04What still fails

honest numbers, 13.10.2026, Instagram Graph API

In the last 30 days @vektor.fm posted 18 films. One passed 10,000 views (27,956). The reels of the last 14 days range from 142 to 3,763 views. This is not a hit machine. Its audience is consumers, so it shows the system running, not buyers coming in.

  • Reach is uneven. The reviewers catch weak scripts, but they cannot predict what people watch.
  • Sign-off is the bottleneck. Five decisions per film cost real time. No current figure for hours per film is published here.
  • Upload and posting are manual by design, so a human always sees the final result in the app.
  • The shared core is not finished. A common engine for both brands, with each brand's data kept sealed, was started on 09.10.2026 and is a scaffold today. Both brands still run from their own folders (status as of 09.10.2026, from the project's README).

05Same system, for your team

The pattern is not about video. It is: split the work into roles, give each agent one job, put a person at the few points that matter, add reviewers that can block, and keep a record of every approval. That is what I set up for teams.

  • Fractional Head of AI. I act as your head of AI two half-days a month: priorities, tool choices, review of what your team builds. See the offer.
  • Agent operations. I build and run a crew like this for one process in your firm, with evals, human sign-off and runbooks. See the offers on the home page.
  • Start smaller. The team AI audit checklist shows where your team stands in 20 minutes.