Forty Percent of the Code, a Quarter of the Staff

Dear Young Comrades,

In the same week, the same company made three separate announcements. Read apart, one looks like a human-resources story, one like a product launch, one like a market story. Put side by side, they make a single picture of where computing labour is being taken today.

• 26 October, Monday: Microsoft sent a "voluntary exit" offer to about a quarter of its workforce, 60,000 employees (CNBC, GeekWire).
• 28 October, Wednesday morning: GitHub, which Microsoft owns, released its annual Octoverse report: in September 2026 more than 40 percent of pull requests opened began with a draft prepared by an AI agent. In March that share was 17 percent (Developer Tech, The Next Web).
• 28 October, Wednesday evening: Microsoft reported $36.6 billion in net income for the July–September quarter, a 32 percent rise on the same period a year earlier (Yahoo Finance).

The three-sentence summary is this: The machine now starts a part of the code; the productivity gain is written into the profit line; the cost is charged to the worker by melting the staff. In this piece I will first summarise that picture briefly, then take each part in detail and try to show the old class relation under the story of "new technology".

---

First, the Summary: Seven Theses

1. Forty percent does not mean "the programmer is finished"; it shows that the content of the labour has changed. The agent starts the draft; the human reads, corrects, approves and takes the responsibility. The labour of writing is turning into the labour of supervision.
2. This transformation is mental Taylorism entering the software department. Conception and execution are being separated; execution goes to the machine, and supervising the machine is left to a human job that is increasingly fragmented and sped up.
3. Everything the agents "know" was learned from the common labour of millions of developers. Octoverse itself is a report of the data produced by 225 million developers. The expropriation of the general intellect here turns into a concrete figure.
4. The offer to 60,000 people is not a "layoff"; it is a layoff dressed as voluntariness. The company defines it as "standard severance"; there is no extra incentive, only the transfer of risk to the worker.
5. Showing the door to a quarter of the staff in a quarter when profit rose 32 percent is a decision of accumulation, not of crisis. Money is moving from the person to the machine, that is, from variable capital to constant capital: $45.2 billion of capital expenditure in the same quarter.
6. "Voluntariness" is the cheapest instrument of discipline applied to the unorganised worker. Each employee is asked, one by one, "do you leave now, or are you pushed later?"; there is no collective bargaining table.
7. The answer cannot be individual either. Turning the productivity gain into a shorter working time, into employment security and into shared commons is possible only through the organised power of computing workers.

Now to the detail.

---

What Do the Figures Say?

GitHub Octoverse 2026: The Agent Takes a Seat

GitHub's report covering 1 October 2025 – 30 September 2026, and the announcements at the GitHub Universe event on 28 October, draw this picture (Developer Tech, The Next Web, Security Boulevard):

Indicator	Value
Pull requests that began with an agent draft (September 2026)	Over 40% (17% in March)
Pull requests agents started in September alone	110 million (about 6 times March)
Agent-originated requests merged within two weeks	From 83% (February) to 90% (September)
The same rate for requests with no agent	About 86% (unchanged)
Share of users of Copilot's agent features	55% (35% a year earlier)
Yearly rise in use of the Copilot cloud agent	230%
Commits made over the year	34.3 billion
Total developers	Over 225 million (US 38.3 million, India 32 million)

Two details need underlining. First, GitHub itself says the agent count is a lower bound, that it cannot detect every agent. So the real share is probably higher. Second, "started by the agent" does not mean "finished by the agent". Every agent draft has to be read, corrected and approved by a human before it is merged. The knot of the matter is exactly in that moment of "approval".

The same day GitHub announced that Copilot was moving from a subscription to a model of paying for the AI credit used. In Satya Nadella's words: "We are redefining the software development lifecycle for the agent era." (The Next Web)

Microsoft's Exit Offer: The Anatomy of Voluntariness

The known details of the offer are these (KIRO 7, Tech Insider, CTV News):

• Scope: About 60,000 employees in the United States, Canada, India and the United Kingdom, at vice-president level and below, in "participating teams".
• Those left out: Sales, marketing and some engineering teams. The core that brings in revenue and that builds the agent systems is being protected.
• Terms: In the company's own words, "not a buyout programme, a standard severance offer". The extra incentives of the early-retirement programme at the start of the year are not here.
• Calendar: The deadline for a decision is the beginning of December; departures run until the end of 2026.
• Expectation: A Microsoft spokesperson says that in similar programmes only 5–6 percent of employees leave. That corresponds to a figure of about 3,000 to 3,600 people. Joseph Philips, an economist at the University of Washington, thinks the company has a target figure and hopes to reach it by a voluntary rather than a compulsory route.

In a note from human-resources director Amy Coleman, the programme is offered as "an important step in how we adapt to change" and as "the option to make this transition now, with support".

This offer is not an event on its own either. The chronology of the last eighteen months is this:

Date	Event
May 2025	About 6,000 compulsory layoffs
July 2025	About 9,000 compulsory layoffs
April 2026	An early-retirement offer to more than 8,600 US employees; about a third accepted
July 2026	About 4,800 compulsory layoffs (2.1% of the workforce)
October 2026	A "voluntary exit" offer to 60,000 people

The Balance Sheet: The Same Week's Profit Line

The July–September 2026 quarter results, announced two days after the offer was sent (Yahoo Finance, Stock Titan):

• Revenue: $92.3 billion (up 19%)
• Net income: $36.6 billion (up 32%)
• Azure cloud growth: 48%
• Quarterly capital expenditure (including finance leases): $45.2 billion
• Returned to shareholders as dividends and buybacks: $10.2 billion
• Contracted commercial revenue still to be billed: $677 billion (up 73%)

So this is not a company in crisis. A company whose profit is breaking records, whose cloud demand outruns supply, and which distributed $10 billion to its shareholders in a single quarter, is asking a quarter of its employees in the same week, "would you like to go?"

---

First Detail: From the Labour of Writing to the Labour of Supervision

What does the agent do, what does the human do?

A pull request is the formal application for a change in software to be joined to the main code. Once, the same person prepared that application, thought the code, wrote it, tested it and explained the reason. In a growing share of cases the process now runs like this:

1. A work record (an issue) or a security alert is assigned to the agent.
2. The agent prepares a draft that changes more than one file (87% of agent-originated requests touch more than one file).
3. The human reads this draft, tries to understand it, corrects its errors.
4. The human approves and takes responsibility for the change.

The most striking thing in the Octoverse figures is that the merge rate of agent-originated requests has risen to 90%, while that of human-originated ones stays at 86%. This figure can be read in two ways. The company's reading: "Agents write better code." The class reading: "The pressure to approve an agent draft is stronger than the freedom to criticise a human draft." These data alone do not say which reading is right. But it does not take a prophecy to know that a reviewer working under the pressure of speed, credit and measurement will grow more inclined to press the "approve" button.

The tools GitHub announced the same day confirm this direction: the "Max" version of Copilot's code review, giving Copilot approval authority on certain file paths, and the ability to assign up to 25 alerts to one pull request in automatic security fixes. Supervision itself is being mechanised; what remains for the human is to sign at the last link of a chain in which the machine supervises the machine.

Mental Taylorism enters the software department

What Frederick Taylor did in the factory a hundred years ago was to take the knowledge in the worker's head and turn it into instruction cards in the hands of management. Conception stayed with management, execution with the worker. Harry Braverman named this, in Labor and Monopoly Capital, "the separation of conception and execution", and defined it as the essence of deskilling.

Computing labour long seemed to stand outside this separation. The programmer both conceived and executed; that also gave them a relatively high wage, flexibility, and the illusion "I am not a worker". We discussed this illusion earlier, in The Tearing of the Illusion, as a profile caught between a "mental proletariat" and a "labour aristocracy".

The agent era carries this separation inside software:

• Conception: A small core that sets the architecture, configures the agents, and chooses which model looks at which job. (The "some engineering teams" Microsoft exempted from the exit offer are very likely exactly this place.)
• Execution: The agents.
• Supervision: A wide mass of reviewers. This labour is less visible than writing, harder to measure, easier to speed up and easier to move.

This third layer does not carry the creative satisfaction of the labour of writing; but it carries the responsibility for the error. If the agent writes wrongly, the person responsible is the human who approved it. The machine produces, the human stands surety. This is not a new job description; it is a new form of an old class relation. We discussed this side from the angle of surveillance in When AI Becomes the Boss's Eye; here we see the production side of the same process.

The credit model: tying labour to a meter

Copilot's move from a subscription to credit-based payment looks like a technical pricing detail. Yet it says something important: the company no longer invoices "how many developers do you have?" but "how many credits did you spend?" The cost item of software production is beginning to be measured per computation, not per person.

For an employer the meaning is simple: a developer's monthly cost and the cost of the agent credit that does the same job can now be compared on the same table, in the same unit. The wage is being raced against a meter value.

---

Second Detail: From Whose Labour Did the Agent Learn?

While reading the Octoverse report one thing should not be forgotten: The report itself is a record of labour. 34.3 billion commits are the record of the work that 225 million developers wrote, corrected, argued over and documented through the year.

That agents can "write code" became possible because knowledge accumulated for decades by open-source communities, students, volunteers and waged developers in public repositories was transferred into models. This accumulation is one of the most concrete present forms of what Marx, in the Grundrisse, called the "general intellect": knowledge produced socially, owned by no one alone.

Once this knowledge has been turned into agents:

• It is sold back to those who produced it in exchange for credit,
• It is used to make the jobs of some of those who produced it look unnecessary,
• And every new step of production is recorded again and becomes the raw material of the next model.

Since The Expropriation of the General Intellect and the AGI Illusion we have called this "the expropriation of the general intellect". In the GitHub and Microsoft example this process is completed under the roof of a single company: the platform that collects common knowledge (GitHub), the infrastructure that turns it into a model (Azure), the product that sells the model (Copilot), and the labour declared unnecessary in this process (the 60,000 people receiving the exit offer) stand side by side on the same balance sheet. The vertical integration of monopoly capitalism cannot be shown more plainly.

---

Third Detail: Why Is Voluntariness an Instrument of Discipline?

Whose "option" is the option?

The phrase in Coleman's note is carefully chosen: employees are offered an "option". But let us look at the content of this option:

• No extra incentive; only legal or standard severance.
• Which teams are "participating" is not announced; the employee learns whether they are a target by receiving the offer.
• The deadline is the beginning of December; the decision is squeezed into the period before the new year, when uncertainty is most intense.
• And three waves of compulsory layoff have been lived through in the last eighteen months.

In these conditions the question in front of an employee who receives a "voluntary exit" offer is in fact this: "Do you go now of your own wish, or in a compulsory wave a few months later?" This is not a choice; it is the imposition of the same result in two different forms. But because the cost of the second to the company (reputation, legal risk, public reaction) is higher, the first is packaged as an "option".

We took up this pattern through another example in No One Was Fired, the Positions Melted Away: a quiet melting instead of a mass layoff. There is no headline, no call to action, no collective point of resistance forms. Each employee decides alone, at their own desk, at the head of their own email.

Voluntariness and the lack of organisation

Voluntary exit programmes work best where the workplace is unorganised. Where there is a union or a collective agreement, such an offer comes to the table not as an individual email but as a subject of collective bargaining: How many people? By what measure? With what compensation? What will the workload of those who remain be? To whom will the productivity gain go?

There is an example of this in Microsoft's own history: while buying game companies, the company signed a neutrality agreement on unionising with the Communications Workers of America, and quality-assurance workers in the affiliated studios unionised and made a collective agreement. So even under the Microsoft roof, organised labour can be seen to have a different bargaining power. Yet the greater part of the white-collar computing staff targeted by the offer to 60,000 people lacks this protection.

We discussed this with the figures from Türkiye in The Misery of the White Collar: to the extent that the white-collar worker does not count themselves a worker, they also shut themselves out of the protections the worker has historically won.

---

Fourth Detail: Where Does the Money Flow From, and To?

From variable capital to constant capital

Marx divides capital into two parts: constant capital, invested in machines, buildings and raw material, and variable capital, paid for labour power. Constant capital only transfers its own value to the product; the only thing that creates new value is living labour. Capitalist competition, though, constantly pushes capital to raise the share of constant capital, that is, towards mechanisation.

Microsoft's July–September quarter is like a textbook example of this law:

Item	Value
Quarterly capital expenditure (data centre, chip, infrastructure)	$45.2 billion
The same quarter a year earlier	$34.9 billion
Investment planned for the 2026 calendar year	About $175 billion
The move toward the workforce in the same period	An exit offer to 60,000 people

Capital is flowing from the person to the machine. But there is a contradiction here: if the only thing that creates new value is living labour, capital that reduces labour and invests in the machine faces, in the long run, the tendency of the rate of profit to fall. Today's 32 percent rise in profit does not refute this tendency; it shows that the monopoly rent provided by a monopoly position, that is, by owning an infrastructure rivals cannot produce, temporarily covers this tendency. We had this discussion through Meta's $145 billion investment in Bubble or Debt?: who wins while it inflates, and who will pay if it bursts, are both clear.

To whom does the productivity gain go?

Let us suppose the agents really do raise productivity. Suppose a developer can now supervise and merge in three days the work they once did in a week. This gain can go to four places:

Where does the productivity gain go?	Who decides?
More work for the same wage (intensification)	The company
The same work with fewer employees (melting the staff)	The company
Dividend and buyback to the shareholder	The company
A shorter working time for the same wage	Only the organised worker can win this

Microsoft's preference is plain in the table: the first three options are applied at the same time. The fourth is the only option no company will choose of itself. The history of the eight-hour day taught us this: as we told in What Did the October Revolution Give the World?, the shortening of working time is not a spontaneous gift of technology; it is a gain of class struggle.

---

Two Languages, Two Readings

The company's language	The class's language
"Agents start 40% of pull requests."	The labour of writing is being turned into a labour of supervision whose responsibility stays with the human.
"The merge rate of agent-originated requests is 90%."	The pressure to approve may be wearing down critical review.
"We are redefining the software development lifecycle."	The separation of conception and execution is being carried into software: mental Taylorism.
"Pay only for the credit you use."	The wage is being raced against a meter value.
"We are offering employees a new option."	Risk is transferred to the individual; no collective bargaining table is set up.
"We are adapting to change."	The staff is being melted while profit breaks records; this is a decision of accumulation, not of crisis.
"Cost control in order to finance AI investments."	Variable capital is transferred to constant capital; living labour is turned into dead labour.
"A community of 225 million developers."	The common labour of 225 million people is the raw material of a single monopoly.

---

Seen from Türkiye

This picture does not concern only Redmond, Bengaluru or London. For the computing worker in Türkiye it has at least three direct results:

First, we are the weakest link in the outsourcing chain. A significant part of software companies in Türkiye does project-based or subcontracted work for customers abroad. The work that agents will displace first is exactly this defined, fragmented work of the "prepare a draft, deliver it" kind. While the company at the centre melts its own staff through voluntariness, the subcontractor's staff at the periphery is melted by a contract not being renewed, without even an offer being made.

Second, the entry step is closing. The work agents take over most easily is the work once given to new graduates: small bugs, dependency updates, writing tests. According to Octoverse, more than 30 million requests from automatic dependency updates alone were merged. The "bottom rung of the ladder" problem in The Misery of the White Collar, where we discussed the fall in white-collar hiring, is lived even more sharply in computing.

Third, the lack of organisation is deeper here. The computing worker in Türkiye is under a triple siege: not being recognised as a separate branch of industry, the barrier of branch no. 10, and the ideology "I am not a worker, I am a professional". While a worker at Microsoft at least has a chance of becoming a story that catches the public's attention, a staff melted in a software house in Istanbul does not even become news. We discussed these barriers, and the ways past them, in detail in The Computing Worker's Handbook.

---

Concrete Tasks

In the face of this picture it is easy to lament; listing what can be done is harder, and more necessary. My proposals for computing workers, the computing union and professional organisations are these:

At the workplace

1. Make the labour of supervision visible. Record the time spent reviewing agent drafts. Labour that is not seen cannot be a subject of bargaining.
2. Ask the question of "approval" responsibility. When an error comes out in code the agent wrote, who is responsible? This question should be written into the job description, the contract and the workplace rules.
3. Do not answer voluntary exit offers alone. Talk with the coworkers who received the offer, compare the conditions, prepare common questions. The moment of an individual decision can be turned into a moment of collective discussion.

At union level

4. Artificial-intelligence clauses in collective agreements. Tying the use of agents to employment security, transferring the productivity gain to a shortening of working time, and consulting worker representatives before new tools are brought in.
5. Limits on telemetry and measurement. Banning the use of agent-use data as a basis for individual performance evaluation and for layoff decisions.
6. The struggle for computing to be recognised as a separate branch of industry. Without this struggle none of the clauses above can settle on a legal ground.

At the level of the commons

7. Making open-source contributions with licence awareness. Discussing licence and data-use terms against common labour being free raw material for monopoly models.
8. The people's data, the people's model. Developing public and cooperative software and model infrastructures against dependence on monopoly agents. We began this discussion in the Socialist AI Manifesto; we will continue it at ÖYD's Odak 2026 conference in December.

At the level of alliance

9. The computing worker is not alone. The same process is at work in the data centres where the code the agents started runs, in the warehouse worker managed by an algorithm, and in the courier managed by an app. No section can win alone without building this ring of alliance.

---

The Machine Is New, the Chain Is Old

Dear young comrades,

GitHub's 40 percent and Microsoft's 60,000 are not two separate stories of the same week; they are two faces of the same process. On one face stands the extraordinary power of technology: the accumulation of millions of people, turned into systems that can produce new code in seconds. On the other face stands the question of in whose hands this power is: the same accumulation is being used to show the door to a quarter of those who produced it.

The technology itself is not guilty. An agent preparing a draft can free a human from repetitive work; it can shorten working time; it can let more people take part in producing software. None of these will be a spontaneous result of the technology. All of them depend on the answer given to the question of in whose hands, under whose control, and for whose benefit this technology works.

The company's human-resources email will not give that answer. The answer given together by the 60,000 people who receive the email, and by their colleagues in Türkiye, in India, and everywhere in the world, will determine it.

They can teach the machine to write code. They cannot teach it to organise.

---