Coinbase's CEO Runs 3 Tiers of Agents. A chief of staff sits on top.
He gave the shape and skipped every number.
Almost every description of an agent workflow comes from somebody selling agent workflows. This one came from the chief executive of a public company, describing his own week, with nothing attached to buy.
Armstrong
"I've changed how I work over the last few weeks. I now have one lead agent for each area of focus, running all the time, and a chief of staff agent that oversees them. Each lead manages its own sub-agents that write code, test it, chase reviews, and coordinate with team members."
Read the structure rather than the enthusiasm. There are three tiers, and each one has a different job.
It is an org chart
At the top, a chief of staff agent whose work is oversight rather than output. Below it, one lead agent per area of focus, and the load-bearing phrase there is running all the time. Not opened, prompted and closed. Standing. Below those, sub-agents doing the work that has a definition of done: writing code, testing it, chasing reviews, coordinating with the humans.
That has a span of control, a delegation boundary and an escalation path. It was not designed by an engineering team and handed upward. It was designed by the person at the top, which is the part worth noticing, because the usual failure of AI inside a company is that the people who pick the tools have no authority over how the work is organised.
The operating rules he left out
Armstrong gave the shape. The next day Hiten Shah quoted him and supplied the part a reader can actually copy: the rules his own team has run this way since February.
Shah
"Your coding agents should have a manager. That manager should be a coding agent."
Shah's version, in five rules:
- One lead agent per project. It holds the plan, remembers the decisions, delegates, and knows what is stuck.
- Task agents work in parallel under it. They take questions back to the lead, not to you, and report results there.
- Agents show the work. Screenshots, live pages, the alternatives they considered, and what changed since the last round.
- One review queue for the human. Everything that needs judgment lands in a single place, where you approve, request changes or give feedback, and the answer routes back to the agent that asked.
- Feedback stays open until it is handled. Changes are tested before they land, merges get a second opinion, and risky changes need approval.
Two days later Anthropic published a run that shows what that layer is for: 1,000 agents went flat at 86% and only finished because somebody restarted them.
Rule four is the one that makes the hierarchy pay. Without it, every agent interrupts you directly and the org chart is decoration. With it, the human stops supervising agents and starts reviewing finished work.
It converges on a hierarchy for the same reason companies did. A single agent handed everything runs out of context and out of attention, and the work it drops is invisible because nothing is watching it. Splitting by area and putting something above it to arbitrate is the cheapest available fix. Organisations of humans found that answer too, and for the same reason.
Copy the design, not the claim
Treat this as a shape to borrow rather than a result to cite. The pattern it names has appeared in this Playbook twice already, from operators rather than executives, and it keeps resolving the same way: something has to sit above the agents, and the thing that sits above them works best when it is also an agent.
What changed on 7 November is who said it. Not a founder of a tools company, not a researcher, not an anonymous thread. The chief executive of a listed company, describing the way he personally works now, in four sentences with an org chart inside them.
So the useful question is not how many agents you are running. It is whether anything sits above them, whether there is one queue where their work reaches you, and whether you could sketch the result on one page.
Draw the org chart.



