We run our companies with agents. Here is what a normal Tuesday looks like

Orwell Lab operates a portfolio of its own businesses: lead-generation publishing sites, financial comparison platforms, software products. The workforce is AI workers organised into departments. The management is one person. People ask what that actually looks like day to day, and the honest answer is that it looks like management, with the interesting parts left in and the drudgery removed.

Morning: reading, not doing

The overnight runs finish before anyone is awake. Research workers have pulled the search and market data. Content workers have drafted. Technical workers have checked the sites, the links and the feeds. By morning the system has organised itself into a queue: here is what was produced, here is what passed automated review, here is what needs a decision.

The first hour of the day is that queue. Approve, adjust, reject with a note. A rejection is not a failure of the system; it is the review layer doing its job, and the note becomes part of what the workers learn from. Most items take seconds because most work is routine and the criteria are clear. The items that take minutes are the reason a person is still in the chair.

Midday: exceptions and flags

A typical flag from the monitoring layer looks like this: a metric on one site has moved outside its usual range, or a worker has failed the same check twice, or a source that a research worker relies on has changed shape. The system's rule is to stop and ask rather than improvise, so a flag arrives with context attached: what happened, what the worker was trying to do, what it recommends.

Some flags take a decision. Some take a change to a worker's brief, which is edited once and applies from the next run. The pattern that took getting used to is that fixing the worker fixes every future Tuesday, which is not how managing people works.

Afternoon: the actual job

With the operation running itself, the afternoon is the work that was always supposed to be the job: deciding what the businesses should do next. Which site gets a new section. Which process gets mapped next. Which experiment gets switched off. The system produces the reports that inform those calls, weekly and on demand, and then it waits to be pointed somewhere.

What it does not look like

It does not look like pressing a button marked "run company". Exceptions still land on a person, because that is the design, not a shortcoming. The workers do not have taste, and they are not asked to; direction, standards and judgement about what matters remain stubbornly human. And when something new is added, it goes through the same shadow-mode probation everything else did.

What has genuinely changed is the ratio. One person's judgement now stretches across an operation that would otherwise need a floor of people, because the judgement is spent only where it is needed. That is the whole promise, delivered in the most boring way possible, one Tuesday at a time.