Multi-agent operating model

What it does

Runs the studio's work through several scoped AI agent lanes, each with its own instructions and memory, while a person keeps every decision that moves money or reaches the outside world. Work counts as done only after a written report passes a check and a separate lane reviews it.

How it works

• Each lane has its own instructions and memory, loaded by a startup hook when the lane starts.
• A written report gate checks that a task states what changed, how it was verified, what could not be verified, time and spend, and what authorised it.
• A separate audit lane reviews other lanes' finished work before it is accepted.
• Human-only gates protect money, publishing, sending, deploying and deleting.
• Cheap-to-expensive model routing sends routine work to cheaper models.

Guardrails and design choices

• Money, publishing, sending, deploying and deleting stay human-only.
• A task must pass the written report gate before it counts as done.
• A separate audit lane reviews other lanes' finished work.

Status

Live since October 2026 and in daily use.