Fork your boss

We have spent two years asking which jobs on the work floor AI will cut. The most-starred thing I saw in August was a management layer you can git clone. On Friday it had 1,797 stars; by Saturday morning, 2,606.

Open Executive is one "executive" persona backed by eight specialist agents, from Chief Strategy Officer to General Counsel, that answers you in Slack and remembers what it advised last month. It reached the Hacker News front page on 27 August under a headline that tells you everything about the mood: "CEO fired developers to make room for AI. Developers create open source AI CEO." That headline is probably folklore; nobody could find a firing anywhere in the repository. The thousand upvotes are not.

Automating the work floor is slow, expensive and, as Meta has just found out, unreliable. Automating the layer above it is cheap, fast and now open source. The only reason we keep attempting the first is that the second would have to be done by the floor itself.

What management does, stripped to its function

A manager collects information from below, analyses it, and passes a version of it upward; then takes a decision from above and passes a version of it down. Each layer repeats the loop with a wider view. That is a language model with retrieval, memory and a scheduler. A Chinese management consultant said it more brutally in August: the middle manager's moat is the information gap, a toll station on a road the manager didn't build.

Now describe a nurse's Tuesday the same way. She notices a resident is "not himself" or "there's something off" two days before any number moves. She decides which of three urgent things is most urgent, standing in a corridor, while a family member asks about a fourth. She holds a licence that makes her personally accountable for what she does with her hands. You can't write that down as a prompt, and nobody has published a repository for it. The tacit, licensed work is expensive to automate; the relay work is cheap.

Meta tried it from the top, and it broke

The first organisation to try it at scale was Meta. A Reuters investigation published on 26 August describes Project OT, hatched at Mark Zuckerberg's January retreat: teams of ten to twenty replaced by pods of three to five "builders", layers of middle management eliminated, one unit head for every 30 to 50 people, "agent-assisted analysis" setting daily priorities, and scenarios that cut some teams by 60%.

It came apart in four months. Code changes were up 220% year on year, while features that reached users were up 36%. Unchecked agents took "large-scale, disruptive actions that humans are unlikely to execute"; major incidents rose 40% and firefighting time rose 70%. Employees, told to install keystroke tracking so the agents could learn their jobs, went into open revolt. Zuckerberg cancelled the second wave of cuts the night before the first one went out.

That isn't evidence against the thesis. Meta did delete management layers, and the pods are still there. What failed was the direction. Management deployed agents onto the floor, tracked the floor to train them, and then discovered that the floor was where the reliability lived. The people who could tell a disruptive action from a normal one were the ones being replaced. Notice where Meta aimed the agents: at the builders' work, not at the reports flowing up to the unit heads. The hard half got four months and a 10% layoff; the easy half got a new org chart and stayed human.

The inversion

What happens when the floor deploys its own management?

Healthcare has the deepest management stack per unit of frontline work. Administrative staff in US healthcare grew by roughly 3,200% between 1975 and 2010 while physicians grew by 150%, and the American Hospital Association now puts administration at more than 40% of hospital expenses. A common response to AI in health systems is to appoint a Chief AI Officer: a new layer to manage the technology that automates layers. This is the empty layer I wrote about in March, with the agency reversed.

The people who could run an "open executive" are already here. A GP practice, a home-nursing team, a physio group: load your protocols, rosters and contracts into it, and let it draft the board pack, chase the unpaid invoices, flag the roster gap before it becomes an agency shift and remind you which contract renews in March. Keep the head nurse and the senior clinicians, and fork the four layers above them. For a group of clinicians who code, that is a weekend, not a moonshot.

What I have to concede

The Hacker News thread holds the steelman. One commenter, a C-suite regular, wrote that "you cannot replace blames, decision in the face of uncertainty". Another: "An auditable executive would certainly be a first." Managers absorb ambiguity, carry consequences, protect their people from the layer above, and decide when the documents don't cover the case. Azeem Azhar made the deeper point in August: agents lack "reputation, recourse and protection for the lone dissenter". And one reviewer's entire verdict was "a good way to run up a really large anthropic token bill".

So a forked management layer still needs one named human who can pull the plug, sign the roster and take the call at 3am. What changes is where that human sits, and it is no longer a corner office.

Which layer of your organisation would you fork first: the one that holds the licence, or the one that holds the meeting?

💥 May this inspire you to audit your own org chart for relay work before someone else does it for you.