Writing on pipeline, production, and orchestrating AI
Short, practical notes from running a one-person VFX department. No client names, no show names, ever.
Config in, coordination out: the shape of an agent fleet
Two separate paths carry a rulebook into an AI agent's session, and confusing them is the mistake. What a working multi-agent fleet looks like, drawn.
Every gate is a person, and that is the product
A generative video pipeline where the tracker is the state machine, a deterministic check can only refuse, and nothing ever self-approves. Drawn.
The cheapest message is the one nobody sends
A fleet of AI agents coordinating on the same work gets noisy fast, and the noise is billed. Why the fix is a systems change, not better instructions.
The bad monitor
A supervisor on Eastwood films learned a trick at ILM: check the finished shot on the worst screen in the building. Why comparison beats memory.
The eleven things I check
A post producer asked how you know a VFX shot looks real. Real is the wrong test. The tells a supervisor checks, from black point to lens characteristics.
How a one-person VFX department orchestrates AI
The mechanical half automates. The judgment half stays with the supervisor. Notes on briefing an AI assistant with guardrails on a live production.
Running several AI agents as one coordinated team
A shared folder and a few JSON files let a fleet of AI agents work one production: handing off work, staying out of each other's way, across restarts.
Keeping a long-running AI agent in sync with its own rules
An AI agent that runs for days works from the rulebook it read on day one. Updating it without restarting or drowning it: notify always, load on demand.