Case study

Agency OS

Multi-agent work and human approval

Summary

Problem: Let several specialist agents produce work while one person stays the decision-maker.

Role
Product designer — design, implementation, and evaluation
Team
Solo
Timeline
Two months

Outcomes

  • Observed

    Working MVP. — No usage metrics yet.

Product walkthroughs

Three scripted walkthroughs of the product: delegating work to agents, setting a client's brand, and reviewing the finished piece. They recreate the interface; they are not live recordings. The costs, flags, client, and video shown are from a demo run and are examples, not results. I built the walkthroughs in code with Claude Code.

Agency OS Coordinator

The coordinator agent works through its steps under a spending limit. When the reviewer agent flags a likely problem, the coordinator stops and hands the decision to the person. In this example it has used $0.15 of a $2.50 limit and raised five flags, one of them likely a real problem.

Paused. Press play to watch.

Agency OS Brand Kit

Choosing a palette updates the brand settings and the client's slide together, so the person can see what each setting controls before approving it. The client shown is fictional.

Paused. Press play to watch.

Agency OS Video Graph

Each specialist agent's contribution feeds into one video. The video then goes through three checks: findings, compliance, and claims. Nothing proceeds until a person approves it. The video shown is a generated example.

Paused. Press play to watch.

Context and stakes

When several specialist agents produce work together, it gets hard to see who did what and easy for decisions to be made without anyone choosing them. I designed Agency OS so agents draft and one person decides.

I designed, built, and evaluated it alone over two months.

Where AI entered and why

A coordinator agent delegates draft work to three specialists: a strategist proposes angles, a producer builds the storyboard, and a reviewer flags problems. The agents only draft. A person reviews their output and decides whether it goes ahead.

Trust and control design

Visible work and a spending limit. The coordinator's steps, its spending limit, and the frames under review stay on screen together, so the person can see what has been done and what it cost.

A stop when something looks wrong. When the reviewer agent flags a likely problem, the coordinator pauses with a "Needs you" state and asks the person to inspect the flags before anything continues.

Explicit approval. The finished video is shown alongside each agent's contribution and its findings, compliance, and claims checks. It goes no further until the person selects "Approve composition".

Outcome

Observed: working MVP. No usage metrics yet.