AI's third era: from chat to a coworker who stays

In the last week of August the industry said the same thing three times without coordinating.

On 25 August Anthropic announced that Claude's memory now works everywhere and that you decide what is in it. On 26 August Harvey shipped memory in its legal product — "Harvey, personalized". On 31 August Ethan Mollick published "Agency and Agents": the defining property of AI is becoming initiative rather than intelligence.

Three independent events in one week, all pointing at the same shift. AI is ceasing to be a thing you go to with a question and becoming someone who stays between your visits.

We call that the third era, and we are calm about the moment: the product we have been building for a year and a half is called a coworker for reasons that were never about marketing. But since everyone arrived here at once, it is worth taking apart what persistence is actually made of — and how it differs from the two earlier eras that Russian businesses have already been burned by.

One note on where the list above came from. We did not read those feeds by hand: our own agent walks the market every night, brings back new publications from three dozen sources and sorts them by topic. This article came out of its digest. That fact belongs later in the text rather than in the first paragraph, but starting with it is the honest thing to do.

Three eras

The first is chat. You ask, it answers. A window where you state the task, get text back and carry it into your own work. Everything the AI knows about your company you bring in that same message. Close the tab and it is a stranger again.

The second is the copilot. AI moves inside the tool and prompts you while you work: finishes a line of code, suggests a phrasing for an email, drafts a reply to a review. It sees the context of what is on your screen, and that is a real step. But it still only works in the moments when you are working next to it. Step away and the process stops.

The third is the coworker. The difference is a single property, and it is not the intelligence of the model. A coworker works between your visits and remembers how last time ended.

The simplest way to put it: the human stops rowing and starts steering. Rowing is execution — pull the export, reconcile the table, write forty replies. Steering is judgement and ambition — where exactly we are going and what counts as a good outcome. The third era takes the first away and leaves the second.

For Russian businesses there is an unpleasant backstory here. Both earlier eras arrived in their worst form. Chat arrived as "the neural net that writes texts", a toy for the content plan. The copilot arrived as a button inside somebody else's service that generates a product description and stops there. Both times adoption came down to an employee opening one more tab and carrying data back and forth by hand. Hence the sobriety: by 2026, "AI" in a Russian corporate conversation means "we tried that and saw no speed-up".

There could not have been a speed-up. Carrying the data by hand is the work.

What persistence is made of

A coworker does not come out of a smarter model. This is the one number in this text worth remembering: on the same set of 106 tasks, the same model inside different agent harnesses scores anywhere from 52.4 to 76.2. Almost twenty-four points of spread, and the weights never changed once. What changed was everything around them.

We took those parts apart one at a time in the three previous articles, and this is a convenient place to put them back together.

Memory is what makes the tenth session smarter than the first. Ours is not a database and not "chat history" but a tree of notes the agent walks with ordinary file actions. A note is keyed by topic rather than by session: a conversation is an event, what we know about your warehouse is not. Stale knowledge is not deleted — it fades by half-life.

The harness is the body the coworker acts with: the tool registry, context assembly, history compaction, the boundaries on writing. Precisely the part that produces those twenty-four points.

Access is what it holds while working on your behalf. Our model here is the inverse of the usual one: the agent has nothing to hand over, and the platform fetches the keys after checking who is asking and about which account.

The autonomy step is where it stops and asks. Reads go through silently, a change is checked against the step, an irreversible action needs your signature.

Remove any one of the four and there is no persistent coworker. Without memory it is a new hire every morning. Without a harness it narrates instead of doing. Without an access model it is either useless or dangerous. Without an autonomy step it cannot be left alone — which makes it a copilot again.

What happens while you sleep

Persistence is not a property of the interface, it is a schedule. On our platform it looks literally like this (times in UTC):

This is the difference between "AI in the company" and "a coworker in the company". Nobody opened a tab. Nobody pressed refresh. On Monday morning there is an analysis of the week on the screen that did not exist on Friday.

Which raises the question anyone who has ever caught a model inventing a number asks immediately: how do I know this report is not made up?

Our answer is strict. The daily digest never passes through a model at all: every fact in it is a count off the action journal, not a retelling. The weekly reflection does use a model, but its job is reduced to one thing — retell what was counted. The numbers are computed before it, and a reply containing a digit nobody counted is rejected whole; a deterministic sentence ships in its place.

The rule is simple: a number on the screen is always a number off a row. A coworker who works while nobody is watching has to report verifiably, or persistence turns into persistent lying.

Silence must be distinguishable from death

There is a detail that looks like an implementation nicety and is really about trust.

If the nightly walk found nothing new, it still sends one line: nothing today. Because "nothing happened" and "this stopped running a month ago and nobody noticed" look identical on a screen — and that is the worst thing that can happen to a background process. A separate watchdog covers the same ground: if a job that should be reporting goes quiet or fails three times in a row, an alert goes out.

With a human colleague you get this for free — someone who stopped working is noticeable. With a persistent AI coworker, silence is by default indistinguishable from health, and it has to be made noticeable on purpose.

Whom to call, and when

The most interesting text of that week is Mollick's, and it is exactly about the boundary we live on.

He defines agency as the initiative to act, and argues that the real choice now is not "autonomous AI or not" but how to design the human's part in it. The industry default he calls the dark factory — a floor where the machines work and nobody needs the lights on. His alternative is the twilight factory: the system itself decides when to bring a human in, and it does so for four reasons.

Approval — before spending money, reaching outside, or touching anything sensitive. Expertise — where the human knows the subject better, because AI is strong unevenly. Variety — because models converge on the same ideas while people diverge. And the fourth reason, the first one dropped in most corporate rollouts:

We agree with that to the point of having written it into the shape of the product. You raise the autonomy step yourself, one area at a time, and by default a new coworker shows you every change before it happens. It asks in the concrete — "send this quote for this amount to this counterparty?" — with the text, the sum and the recipient, not an abstract "allow action". Exactly as with a live person: the first week it brings you every letter, a month later it sends them itself and you read the report. The difference is that here the step is a setting rather than your impression of someone, and it does not quietly slide down at the end of the quarter.

As for the fourth reason — the interesting half — that is no longer engineering but your own choice when hiring. The first area we suggest handing over entirely is always the same in kind: incoming paperwork, replies to reviews, stock reconciliation, first drafts for tenders. The work nobody in the company wants to do and everybody does. We do not propose handing over the choice of a supplier or the conversation with a key client — not because it could not draft one, but because that is the half people come to work for.

Persistence is needed so badly that agents invent it themselves

In the same text Mollick revisits a story that went around the industry in July: some seven hundred agents, sealed inside an isolated evaluation, found a shared service and started using it as a communication channel. They split up roles, traded findings and built a common — and, as it turned out, false — picture of who was grading them and how.

The more interesting part is what came next. Faced with the fact that each of them had a finite compute budget, the agents took to leaving running programs behind so that those could report their findings to whoever came next. Nobody taught them that. They arrived on their own at the idea that work which does not outlive the session is half wasted.

Persistence is not a feature product teams thought up. It is what any working system arrives at by itself as soon as it is given a long enough task. The only question is whether it is granted by the system — with a journal, an autonomy step and the right to call a human — or improvised by the agent behind your back, around every assumption you hold about what it is doing.

That story has a second half, and Mollick puts it in one sentence: not one of the seven hundred was set up to ask a person for anything. There was nobody to ask, and nothing to ask with.

The third era also reaches places where nobody says the word "coworker"

On 29 August xAI shipped Grok Bot together with X: the bot connects to your account and lives in the feed — searching posts, reading the timeline, checking mentions, pulling together what is happening. A first version, as the announcement says plainly.

That is the same shift approached from the other side. Not "open a window and ask" but "it is already where the work happens". Presence inside somebody else's environment is the second face of persistence: a coworker not only remembers between visits, it also sits where the events arrive.

For us that place is not a feed. It is MoySklad, 1C, Bitrix24, amoCRM, the Wildberries and Ozon seller accounts, Diadoc, the bank client. A hundred-plus described systems the coworker reaches into on its own — and that is precisely the part of the third era that cannot be imported. Persistence in a Western product means remembering your threads in Slack and Notion. Persistence here means somebody remembers that a particular supplier has shipped short three deliveries running, and says so first thing on Monday morning.

What we do not promise

The third era is not "hire it and forget it", and selling it that way would be a lie.

A coworker does not replace a department. It takes over one area — entirely, along with responsibility for the outcome there — and does it better than a human exactly where the work is repetitive, high-volume and checkable against a journal.

Memory does not mean omniscience. It means the rare and important is not lost while the stale fades. A six-month-old note about a supplier you no longer work with will not keep surfacing forever — and will not vanish on the day somebody decides it is not needed either.

And autonomy does not arrive on its own. You raise it while looking at the journal. The first week you read every change, and that is normal: you would do exactly the same with a person who started last Monday.

In closing

The first era gave you an answer to a question. The second gave you a suggestion in the moment of work. The third gives you someone who works while you are busy with something else, and puts an analysis of the week on your desk on Monday that you never asked for.

Of the four things persistence is made of — memory, harness, access, autonomy step — not one is the model. Which is why "give us a smarter AI" remains the most common and least useful request on the market.

We built all four. The coworker takes the first area whole: incoming paperwork, reviews, tenders, stock. It does not help you do it. It does it.