Sort your team's week into work, vision and judgement
- ai
- work
- judgement
Subscribe
New essays and episodes, sent when there is something worth reading. No noise, unsubscribe anytime.
Subscribe
New essays and episodes, sent when there is something worth reading. No noise, unsubscribe anytime.
Twelve hours.
That is roughly where METR put the best publicly available AI models in early 2026. METR measures this as a "50% time horizon": the length of task, counted by how long a human expert takes, that a model completes successfully half the time. Its Frontier Risk Report gave about 12 hours, with a range of 5 to 61 hours. METR also says its task suite can't reliably measure horizons above 16 hours, so read the top of that range loosely. A year earlier, METR had reported that the length of task agents can complete with 50% reliability had been doubling about every seven months for six years.
The tasks behind the figure are primarily software engineering, machine learning and cybersecurity. METR's FAQ asks whether a 2-hour time horizon means an AI can do all the intellectual tasks that a low-context human can do in 2 hours, and the answer starts with "No." The same report cautions that agents' real-world performance has tended to be weaker than a naive reading of their benchmark scores suggests.
So don't take the number literally. What it does show is the direction: agents keep finishing longer and longer tasks, at least in software work. If that carries over to other kinds of work, more of the doing can be handed to them and the doing gets cheaper, though METR measures what models can complete and not what it costs.
Microsoft's 2026 Work Trend Index states the hopeful version of what happens next. As agents take on more of the execution, people get "more room to direct the work, make the calls, and own the outcomes." The line comes from a report built on a survey of 20,000 knowledge workers across 10 markets who already use AI at work.
It's a fine sentence. But it doesn't tell a team what to do differently. Which parts of this week were directing? Which were calls? And which were just the work?
The future of human work is not to do the work. It is to have the vision to direct work and the judgement to make critical decisions.
Every team's week is three piles: work, vision and judgement. If your team has never sorted the piles, they live mixed together in the same calendar, the same task list, the same heads. Teams can prepare for what is coming by pulling them apart, then capturing the vision and the judgement as context: written down, where an agent can read them.
I'm talking about knowledge work here, and only that.
Vision is what you can state in advance: who the work is for, and what should be true if it works.
Judgement is the call that can only be made when the case arrives. You can't write it ahead of time, because the case hasn't shown up yet.
Work is everything else. It's the doing that sits between a stated vision and the next call.
Why does vision stay with people? Because an agent cannot work out from the task itself who the work is for, or what should be true if it succeeds. Nothing in "build the dashboard" says who will open it. So a person has to say. In "Continue" is a complete prompt, I wrote about the files that say who a build is for and what should be true if it works. That was the vision pile for one person's build. This post is about a team's whole week.
In a 2018 working paper, economists Ajay Agrawal, Joshua Gans and Avi Goldfarb treat AI as better prediction. Then they name the separate step a decision still needs, working out its payoffs, meaning what each possible outcome is worth: "We label this process of determining the payoffs 'judgment.'"
Their conclusion is that when prediction gets cheap, judgment becomes more valuable. It comes with a condition: prediction and judgment make each other more valuable only as long as judgment is not too difficult. Their conclusion is about decisions that use AI prediction, and I'm not stretching it to cover all work.
Henry Mintzberg watched five managers for a 1971 study and found work characterised by variety, discontinuity and brevity. "Managers come to prefer issues that are current, specific, and ad hoc, and that are presented in verbal form." He also noted that most of the manager's information was never documented.
Mintzberg was describing information, not vision or judgement. But if most of a manager's information is never documented, then who the work is for, and how the borderline calls get made, are likely to live in heads too.
A person in the room picks that up. An agent reads only what is written.
That is why sorting takes real effort: the vision and the judgement have to be pulled out of people's heads. And sorting alone isn't enough, because an agent can't read what stays in heads. It has to end in context: the vision and the judgement, on a page an agent can read.
This team does not exist. I made it up.
Picture five people who produce a monthly pricing review for an invented retailer. They list everything in their week, then sort it.
Pulling the data, reconciling it, drafting the pack and formatting the charts all go in the work pile.
Who is the review for, and what should be true if it works? Say the answer is that the category managers can act on it without a meeting. That's the vision pile. If nobody has ever written it down, it exists only in people's heads.
Then this month's odd number shows up. Is it a data fault or a real shift? Do they hold the pack? Those calls can only be made when the case arrives, so they go in the judgement pile.
The sort also shows that some items that looked like judgement are really legwork. Checking whether the odd number matches last year's figures turns out to be legwork; only deciding whether to hold the pack is a call. In Automating the slack, I called this "legwork wearing judgment's clothes."
Now the team writes things down. Vision gets a short page. Judgement gets decision records, the decided borderline cases, and a list of what the team's agent never does. It does not get a written decision policy, because the records only hold calls already made, and the next call can only be made when the case arrives.
I'm borrowing the idea from software. Michael Nygard made the case for architecture decision records on software projects in 2011: "One of the hardest things to track during the life of a project is the motivation behind certain decisions." Without it, a newcomer can only accept a decision blindly or change it blindly. His fix is a short record of context, decision, status and consequences.
I believe writing them down helps, and it creates alignment in the team.
But capture has a limit. Harry Collins divides tacit knowledge into three kinds, and only one, relational tacit knowledge, is described as something that could be written down in principle if someone made the effort. The other two are not described that way.
And the piles are not fixed. Items move between them as the tools improve.
A 2000 meta-analysis by William Grove and colleagues found that mechanical prediction, meaning formal statistical methods, was about 10% more accurate on average than clinicians' predictions. That suggests some of what a team files under judgement is really prediction, and where a model predicts better than the person does, the item belongs in the work pile.
Freestyle chess, where human players can use computers during the game, shows the same drift. Garry Kasparov's lesson from a 2005 tournament was that "Weak human + machine + better process was superior to a strong computer alone." Yet a later analysis of the 2005 to 2008 PAL/CSS freestyle tournaments found that once computer-only games from the following five years were included, the freestyle teams kept their lead on the move-matching measure but went toward parity on the others. In 2024, Tyler Cowen wrote that the human's main role in centaur chess was choosing which program to use, and added: "For years now, the engines have been so strong that strategy no longer made sense."
So the sort has to be redone as the tools change.
One: take a real week and sort every item into work, vision or judgement. Argue about the borderline ones, and don't rush that part. Arguing over where those items belong is how the team comes to agree on what its work, vision and judgement are.
Two: write the vision down where the agents can read it. Who is it for? What should be true if it works?
Three: start the judgement record. Log each call as it is made, with its reason. Add the decided borderline cases and the list of what the agent never does.
Keep all of it short. Anthropic's guidance on context engineering treats context as a finite resource with diminishing marginal returns, so the aim is the fewest words that carry the most useful information, or in Anthropic's terms "the smallest possible set of high-signal tokens."
In Nobody demos a text file, I argued that when a team improves the shared files its agents read, every agent it runs afterwards inherits the improvement. The vision page and the judgement record work the same way.
Leaving a human on the sign-off line is not the same as keeping judgement with people.
Raja Parasuraman and Dietrich Manzey's 2010 review concluded that automation bias, the habit of relying too much on what an automated aid says, leads people using imperfect decision aids into errors of both omission and commission. They miss things, and they do the wrong things.
Madeleine Clare Elish named the worse case in a 2019 paper: the "moral crumple zone," where a human operator with limited control over an automated system absorbs the blame for its failures.
So ask one question of every item in the judgement pile. Does the person making this call have the context and the authority to make it differently?
If not, the person is only signing off on a call the system has already made.
Whatever the 12-hour figure turns out to mean for your kind of work, the tasks agents take over will come from the work pile.
The vision pile, and the calls that still need a person, stay with the team. And they are worth more written down than held in heads, because when more of the work moves, the vision and the judgement are already on the page.
Open last week's calendar and start sorting.