In my last post I said the job is moving from doing tasks to owning outcomes. A few days ago I read a story that makes that point better than I did.
The news
On 27 October, Anthropic announced that it had acquired Mithril Medicine, a company with one full-time employee. It works on two rare forms of muscular dystrophy. The price was not disclosed.
The one employee is Cody Tranbarger. He did not run the lab with his own hands. He directed 11 outside laboratories, a few part-time scientific consultants and eight paid advisers, and he used Claude to handle the contracts, the logistics, the experiments and the data that moved between all of them.
In one example from Anthropic’s technical report, as R&D World reports it, Claude wrote a 341-page production report on 113 antibody constructs in under eight minutes. A person on the team overruled the plan, and Claude redid it for 89 constructs in four. Anthropic says turning a programme idea into a design plan took two to three days with Claude, against one to two months without it.
The result so far: proof of principle in mice in under 18 months. If the candidates hold up, human trials could start as soon as late 2028.
What it does not prove
Let me be fair to the story first, because it is easy to oversell.
This is one full-time person, not zero people. The antibodies, the assays and the mouse studies were done by outside labs. Claude also made mistakes. At least three reached outside labs inside emails that people had edited and sent. Two errors in Claude’s review of production data went unnoticed for more than four days.
The science is early, it is in mice, it comes from the company that bought his company, and the report is not peer reviewed. The mouse study was small: seven treated mice against four controls, not randomised, tested months apart.
So I am not telling you AI is curing a disease. I am telling you something smaller and, to me, more important. One person can now carry the coordination that used to need an entire company.
What the one person actually did
Look at what he did all day. He chose which programme to pursue. He decided which lab did which job. He read what came back and decided what to do next. He took responsibility for whether the thing worked.
That is ownership and direction. The operational work, the paperwork and the repetitive chasing went to agents. And the checking stayed human, which is also a kind of ownership. If nobody owns the outcome, nobody catches the error in the data.
It should always have been this way
Here is my case. This is how work should have been organised from the beginning.
Think of the assembly line. Henry Ford’s version split a car into many small steps, and each worker did one step, again and again. It was a brilliant way to build cars. Offices copied the idea. A few people at the top decided the direction and the strategy. Everyone else did a step of the process and handed it on.
It was not done because someone thought people were small. It was done because it was economical. Execution needed human hands, and coordinating people any other way cost too much. So most of us were trained to be good at a step. Do the task well, hand it over, do not ask what the whole thing is for.
That economy is changing. When agents can do the steps, nobody needs a person standing at a station. What is left is the part the assembly line kept for a few people at the top: deciding where to go, and owning whether you got there.
Now everyone is expected to do that. And I think that is where the value of a human will be, very soon. Not in doing the steps well. In owning the result.
Two kinds of ownership
Here is the part that complicates my own argument.
Mithril was built on one person owning an outcome. And it now belongs to the largest AI lab in the world. The very thing I am saying people should do got bought up, in a way that looks a lot like ownership concentrating at the top.
I do not think this kills the point. It sharpens it. Owning the outcome is where the value is going. The open question is who ends up owning the owners. Tranbarger will lead Anthropic’s drug programmes, so he is still the one deciding direction. But the company around him is not his anymore. Ownership as a job and ownership as equity are not the same thing, and the next decade will be about which one people actually get.
A big change, a necessary one, and a painful one
This is a big change. I think it is a necessary one. And I think it will be painful.
Painful because ownership is a muscle, and most of us never exercised it. We were taught to do a step well and pass it on. Owning an outcome means making calls without being told, living with the result, and fixing it when it breaks. That is uncomfortable if you have spent twenty years being good at a step.
It will also be fast. The shift off the farms took most of a century. This one could fit inside a single career. As I said last time, the people in the gap usually pay for the wait.
Humans who own, agents who operate
Here is what I believe works, and what I try to do myself.
Partner with other humans who are willing to own outcomes and deliver them. Look for people who want the result and will take the weight of it, not people who want a list of tasks. Two or three people like that, with clear ownership between them, can do what a department used to do.
Employ AI for the operational and repeat tasks. Everything that is a step, an update, a draft, a chase, a first pass. Give it to the agents, and give it freely.
I do this myself. I run three agents, on three different devices, each with full user access to the systems it works on. That sounds reckless, and without design it would be. But AI has reached a stage where this is workable, if you build the system to be fault tolerant and put fail safes in place. Then the blast radius of an agent can be contained. IT companies have been doing this to their employees for decades, and the same principles now translate to agents. Each of my agents has user and identity management set up, which decides what it can and cannot do, and what it has access to. You do not need to reinvent the wheel. Mistakes will still happen, as the Mithril errors show. The design question is how far one mistake can travel before something stops it.
And hold on to three things that you do not hand over: the decision about what to do, the accountability for how it turns out, and the final check before it goes out. The Mithril story shows why. The agents were fast and very good, and they were still wrong sometimes. Somebody had to own that.
You need to start today
Do not wait for this to arrive at your desk. Pick one outcome you could own, not a task but a result you would be willing to answer for. Then pick one task around it that you could hand to an agent. Do both today, and keep doing them.
It has begun. The only choice left is whether to be the one standing at the station, or the one who owns what the line is for.