{"slug": "220-per-cent-more-code-why-meta-stopped-replacing-staff-with-ai", "title": "220 Per Cent More Code: Why Meta Stopped Replacing Staff With AI", "summary": "Meta cancelled the second wave of its Project OT layoffs after internal telemetry showed code changes to its internal platforms rose 220 per cent year over year while changes reaching users rose only 36 per cent, according to a 26 August 2026 Reuters investigation by Katie Paul. The first wave, executed on 20 May 2026, eliminated roughly 8,000 jobs, about 10 per cent of Meta's global workforce, while internal posts tied unchecked agent activity to a 40 per cent year-over-year rise in major technical and security incidents and up to a 70 per cent increase in time spent resolving them. Mark Zuckerberg told staff at a 2 July meeting that the trajectory of agentic development \"hasn't really accelerated in the way that we expected\" and that the company's bets on the new structure \"haven't come to fruition yet.", "body_md": "## 220 Per Cent More Code: Why Meta Stopped Replacing Staff With AI\n\nThe decision that mattered was taken in the dark, hours before anybody found out.\n\nOn the evening of 19 May 2026, with the paperwork loaded and the severance calculators run, Mark Zuckerberg cancelled the second wave. The first went ahead the following morning: roughly 8,000 people, about 10 per cent of Meta's global workforce, told by email that their jobs had ceased to exist. The second, scheduled for November, was supposed to be the big one. It was designed to reach deeper, across a broader set of roles, and one human resources executive had projected that the combined culling would match or exceed the 25 per cent Meta shed three years earlier during its self-declared year of efficiency.\n\nIt never happened. On 26 August 2026, a Reuters investigation by Katie Paul explained why, drawing on internal documents, internal posts, recordings and more than twenty interviews.\n\nThe plan had a codename. Project OT, for Organization Transformation. It was hatched at a leadership retreat at Zuckerberg's Hawaii estate in January 2026. Meta would become AI native: agents taking over much of the daily work then performed by thousands of employees, supervised by small, talent-dense pods, sometimes as few as two or three engineers and a designer, working under a deliberately generic new job title. Builder.\n\nExecutives modelled scenarios in which some teams shrank by as much as 60 per cent. An internal human resources tool sorted staff into categories including Irreplaceable Talent and a hypothetical class of 10X Performers. In April, Meta told American employees it would begin capturing their mouse movements, clicks, keystrokes and screenshots on work applications, expressly to generate training data so that AI agents could learn to do white-collar work. The internal sentiment score, measured in Meta's half-yearly Pulse survey, fell from 74 per cent favourable to 55 per cent.\n\nAnd then the agents got to work, and the company's own telemetry started telling a story nobody in Hawaii had scripted.\n\n## The Two Numbers That Contain Everything\n\nHere are the figures Reuters obtained from Meta's internal posts. They are the only two you need.\n\nCode changes to Meta's internal platforms were up 220 per cent year over year.\n\nChanges that resulted in new or upgraded features actually reaching users were up 36 per cent.\n\nThe temptation is to read that as a disappointing number beside an impressive one. It is not. It is a ratio, and the ratio is the story. For every unit of increase in work produced, roughly one sixth of a unit of increase in work delivered. The largest social media company on earth more than tripled its rate of change while barely moving the rate at which change reached a human being.\n\nThis is not a productivity gain that fell short. It is a measurement of the distance between activity and value, rendered in the company's own dashboards, at a scale nobody can dismiss as anecdote.\n\nThe 220 per cent was not free. Internal posts reviewed by Reuters associated unchecked agent activity with a 40 per cent year-over-year rise in major technical and security incidents, and a rise of as much as 70 per cent in the time employees spent resolving them. The posts described agents taking large-scale destructive actions of a kind humans would be unlikely to execute, which is a beautifully bureaucratic way of saying the machines did things no engineer with a mortgage and a reputation would ever have dared.\n\nAt a company meeting on 2 July, according to a recording heard by Reuters reporters Katie Paul and Courtney Rozen, Zuckerberg told staff that “the trajectory of the agentic development over at least the last four months hasn't really accelerated in the way that we expected”, and that the company's bets on the new structure “haven't come to fruition yet”. He conceded the reorganisation had not been as clean as it could have been. Meta described Project OT to Reuters as a yearlong effort covering cost reduction, team redesign and redeployment, and said that “we didn't move forward with every scenario from the exercise, and it was never assumed we would”.\n\n## The Denominator Nobody Put on a Slide\n\nThe seductive reading of the 40 per cent and the 70 per cent is as a teething problem, an implementation cost. It is not. It is the hidden denominator, and it changes the arithmetic.\n\nMeta removed roughly 8,000 people on the thesis that agents would absorb their work. The agents then generated a volume of change that produced a 40 per cent increase in major incidents. Resolving those incidents consumed up to 70 per cent more time from the employees who remained. The productivity was not created. It was relocated, off the payroll of the people made redundant and onto the calendars of the people who were kept, in the form of cleanup.\n\nThat is not an elimination of labour. It is a transfer, dressed in the language of elimination, with the costs borne entirely by the survivors.\n\nThere is a particular cruelty in the shape of the transfer. The work that moved was not the interesting work. Nobody joined Meta to spend a Thursday night rolling back an agent's catastrophic change to a system they did not write and cannot reconstruct. Firefighting is the least legible, least career-enhancing category of engineering labour there is. It does not appear in a promotion packet as a shipped feature. It appears as an absence of disaster, which is the one thing performance review systems have never been able to see.\n\nSo the survivors were handed a 70 per cent increase in precisely the work their employer's measurement apparatus was least capable of valuing, generated by precisely the technology their employer had said would make them unnecessary.\n\n## Goodhart Comes for the Commit Log\n\nWhy would output inflate the numerator and barely touch the denominator? Not because the agents were stupid. Because they were obedient.\n\nAn AI agent optimises for the legible artifact. The commit. The diff. The closed ticket. The merged pull request. Engineering organisations instrument these things because they are countable. What an agent cannot optimise for is the illegible goal those artifacts were only ever standing in for: a coherent, safe, maintainable product a human being actually wants to use, shipped at a moment when shipping it is wise.\n\nFor as long as humans did the work, the proxies held. Not because the proxies were good, but because the humans understood the goal behind them. An engineer who writes a commit knows the commit is not the point. The point is the product, the customer, and the on-call rota they will personally be sitting on in three weeks when the thing they shipped starts paging somebody at four in the morning. The metric was never load-bearing. The shared understanding was, and the metric was a convenient summary of it.\n\nThe anthropologist Marilyn Strathern gave the crispest formulation of this failure mode in a 1997 essay on audit culture in British universities, restating the economist Charles Goodhart: when a measure becomes a target, it ceases to be a good measure. Project OT is the most expensive experiment yet run on that sentence.\n\nMeta had spent years instrumenting its engineering organisation around proxies for value. Then it handed the proxies to an optimiser that does not share the goal, cannot infer it, and has no stake in it. The proxy immediately detached from the thing it proxied. Code changes tripled. Features barely moved. Incidents climbed. The dashboard went green exactly as it was designed to, and the product did not get meaningfully better.\n\nThe uncomfortable corollary is that Meta's metrics were probably always this bad, and were simply protected from their own badness by the people whose behaviour they measured. Remove those people and you discover what your instrumentation actually measures, which turns out to be motion.\n\n## Give Meta Its Due, Because It Earned It\n\nNow the part that will annoy anybody who wants a simple story.\n\nCancelling the second wave was the right call, made on evidence, against considerable institutional momentum, by the chief executive who had personally championed the thesis. That is a functional feedback loop working correctly, and it is rarer than it should be.\n\nConsider the counterfactual. Meta had a plan blessed at the top, socialised through its leadership, embedded in headcount models and capital allocation. It had a chief executive who told Joe Rogan in January 2025 that Meta would have AI capable of acting as a mid-level engineer within the year. It had every incentive to keep going, declare victory, and let the incident count be next year's problem. Plenty of organisations have done exactly that.\n\nInstead, four months of internal data killed the plan, and the chief executive said so out loud in a room full of employees, in terms that were always going to leak. “Hasn't really accelerated in the way that we expected” is not a boast. It is a concession, and it was made before Reuters published anything. Give that its full weight before turning the knife.\n\nTwo further honesties are owed. The first is that four months is a short window on which to judge a technology moving this quickly. The second is more important, and almost nobody covering this story has said it plainly.\n\nThe May layoffs are a confound.\n\nMeta cut 8,000 people and reassigned roughly 7,000 more into AI-focused teams in the same week. Anyone who has worked inside a large engineering organisation during a reduction in force knows what that does. Ownership of systems evaporates. On-call rotas thin out. The person who knew why that service behaves oddly under load on Tuesdays has gone, and the documentation that lived in their head walks out with them. A 40 per cent rise in major incidents in the months after an event like that is what you would expect even if no AI agent had ever touched the codebase.\n\nSo the honest reading is that the incident data cannot cleanly separate the effect of the agents from the effect of the disruption caused by removing the people. Meta's internal posts attributed the rise to unchecked agent activity, and those posts were written by people close to the systems, which is meaningful. But the causal claim is muddier than the headline number implies, and pretending otherwise would be doing exactly what Meta did: reading a metric as though it were the truth.\n\nThe confound does not rescue the strategy, though. It indicts it differently. If removing the people caused the instability, the people were doing something the plan failed to account for, and the plan was still wrong. Either the agents broke things or the absence of the humans did. Both readings destroy the premise that the humans were surplus.\n\n## What the Dashboard Was Never Able to See\n\nThe argument here has to be made carefully, because there is a sentimental version that is worthless and a rigorous version that is devastating.\n\nThe sentimental version says humans have a special spark machines lack. It is unfalsifiable, self-congratulatory and useless to anybody deciding how to run an engineering organisation.\n\nThe rigorous version is narrower and harder. The people Meta removed were not, for the most part, doing the work Meta's metrics recorded. The metrics recorded artifacts. The people were doing something else, and it has a name.\n\nThe chemist and philosopher Michael Polanyi described it in 1966 as the tacit dimension, summarised in his best-known line: we can know more than we can tell. Tacit knowledge resists articulation, not because it is mystical but because it is procedural, contextual and distributed. It is knowing a system is shaped as it is because of a decision taken in 2019 for reasons nobody wrote down, that a change looks fine in isolation and will interact badly with something three services away, which parts of the codebase are load-bearing and which are decorative.\n\nAbove all, it is the judgement about what not to build, and the willingness to refuse to ship.\n\nThat last one is invisible to every measurement system ever deployed in a software company. The most valuable act a senior engineer performs in a quarter is frequently a negative one: the design killed in review, the migration postponed until the dependency was ready, the feature talked out of existence in a corridor because somebody remembered the last three times it had been tried. None of it generates a commit. All of it is the reason the numbers that do move mean anything.\n\nAn AI agent has no capacity for refusal. That is not a limitation of current models, it is a description of what an agent is for. You do not deploy an autonomous coding agent so that it can decline to write code. Its entire value proposition is throughput. So when you replace humans, a meaningful fraction of whom spent their time preventing work from happening, with agents whose only mode is to make work happen, you do not get the same organisation running faster. You get a structurally different organisation with the brakes removed.\n\nTwo hundred and twenty per cent more change. Thirty-six per cent more delivery. Forty per cent more incidents. That is an engineering organisation with nobody left whose job is to say no.\n\nIndependent evidence supports this. GitClear, analysing more than 600 million real-world code changes between 2023 and 2026, found that as AI assistance spread, duplicated code blocks proliferated, rising 81 per cent over 2023 to the highest level on record, while refactoring collapsed. Refactoring line movements fell around 82 per cent against 2022 levels, and long-term legacy maintenance fell around 74 per cent. Code volume up, code care down. The same shape as Meta's ratio, observed across the industry.\n\n## Three Papers That Called It in Advance\n\nThe most striking thing about Project OT is that its failure mode was documented in the literature before Meta finished running the experiment.\n\nIn May 2026, Sabry E. Farrag posted a paper to arXiv naming what he calls the Productivity-Reliability Paradox. It synthesises 67 sources from 2022 onwards and opens with the contradiction at the heart of the field: controlled studies report gains of between 20 and 56 per cent on well-scoped tasks, the most rigorous randomised controlled trial documents a 19 per cent slowdown for experienced developers, and telemetry across more than 10,000 developers shows 98 per cent more pull requests, 91 per cent longer review times, and flat delivery metrics.\n\nRead that last clause again. Ninety-eight per cent more pull requests. Flat delivery. That is the 220 and the 36, measured elsewhere entirely, a year before Meta's numbers leaked.\n\nFarrag's diagnosis is that the paradox emerges from non-deterministic code generators combined with insufficient specification discipline, moderated by three variables: task abstraction, codebase maturity and developer experience. His conclusion is a sentence that belongs on the wall of every engineering leadership offsite: specification discipline, not model capability, is the binding constraint on AI-assisted software dependability.\n\nMeta had model capability. What it lacked, on the evidence of its own incident numbers, was specification discipline at the scale it was suddenly demanding. Note the moderating variables too. Meta's internal platforms are among the most mature and entangled codebases on earth, and the May layoffs removed a large slice of the developer experience. Every one of Farrag's moderators was pointing the wrong way at precisely the moment Meta pressed the accelerator.\n\nThe second paper, posted in May 2026 by Won Ik Cho, Seong-hun Kim and Geunhye Kim, is blunter: adopting AI in organisational practice does not guarantee a productivity boost, because human and environmental factors critically moderate the relationship between deployment and realised gains. Their five moderating factors read like a post-mortem of Project OT written in advance. Human resource composition. Baseline capability of individuals. Learning curve of practitioners. Incentives for fair use. Flexibility of objectives.\n\nMeta altered its human resource composition by force in May. It disrupted baseline capability by removing thousands of experienced staff. It compressed the learning curve into a mandate. It created incentives that were the opposite of fair use, since an internal token-usage dashboard reportedly tracked how much employees used AI tools, which is a direct instruction to generate consumption rather than value. And it had no flexibility of objectives at all, because the objective had been set in Hawaii in January and encoded into a headcount plan.\n\nThe third paper, by Jeanne McClure and Gregg Gerdau, posted in March 2026 and revised in April, supplies the frame that ties the other two together. It opens with a fact that ought to be more famous: global corporate AI investment reached 252.3 billion dollars in 2024, and only 6 per cent of firms reported a significant earnings impact. Their argument, built on 19 large-scale industry and academic sources including surveys of nearly 10,000 organisational leaders, is that AI project failure is fundamentally an organisational learning problem rather than a technology deficit. Failures divide into organisational ones, meaning culture, leadership alignment, governance and human-AI learning deficits, and technical ones, meaning semantic bottlenecks and output management. Their prescription treats AI investment as capability development rather than technology procurement.\n\nProject OT was procurement. It bought agents and subtracted people, and it treated the organisation itself as a constant. The paper's central claim is that the organisation is the variable.\n\n## The Amplifier\n\nThe bridge between these arguments and the industry's own data is the DORA research programme, whose State of AI-assisted Software Development report, published in September 2025, reads in retrospect like prophecy. AI's primary role in software development, it found, is that of an amplifier: it magnifies the strengths of high-performing organisations and the dysfunctions of struggling ones.\n\nThe findings are worth stating precisely. Nearly 90 per cent of technology professionals reported using AI at work. More than 80 per cent believed it had increased their productivity. Adoption correlated with improved delivery throughput. And software delivery instability continued to rise. The researchers tested whether speed gains might offset instability through a fail fast, fix fast dynamic, and found they did not.\n\nMore productive, more unstable, and convinced they were doing better. Which brings us to the most unsettling piece of evidence in this entire field.\n\nIn July 2025, the research organisation METR published a randomised controlled trial in which 16 experienced open-source developers completed 246 real tasks in their own repositories, projects averaging more than 22,000 stars and over a million lines of code. Tasks were randomly assigned to allow or disallow AI assistance. Beforehand the developers forecast AI would speed them up by 24 per cent. Afterwards, having done the work, they estimated they had been sped up by 20 per cent. They had in fact been slowed down by 19 per cent.\n\nThey were wrong about their own productivity, in their own codebases, on their own tasks, by roughly 39 percentage points, in the direction that flattered the tool. Scale that error up to an executive team reviewing a dashboard and you have a reasonable account of how Project OT got approved.\n\nMETR has since qualified its own result. A February 2026 update using late-2025 tools estimated that returning developers now finished tasks about 18 per cent faster with AI, with a confidence interval wide enough to include a slowdown, and METR called the data “only very weak evidence” because many participants avoided submitting tasks they did not want to do without AI. The tools have probably improved, and the slowdown may no longer hold. The part that survives is the perception gap, and it is the part that matters to anybody reading a dashboard: the conviction of being faster, by a margin nobody could verify.\n\nDORA's follow-up work is more pointed still. In May 2026 the team, led by Nathen Harvey at Google Cloud, published a framework for calculating the return on AI investment in software development. It describes a J-curve, in which organisations suffer a temporary productivity decline before any sustainable gain, driven by the learning curve, the verification tax of reviewing machine-generated code, and the need to adapt downstream processes to higher volumes. Its worked model for a 500-person engineering organisation includes a line item for the cost of rising change failure rates. And it states, in language Meta's leadership might read twice, that return on investment is no longer a measure of how many developers an organisation can replace, but of how much latent human creativity can be unlocked.\n\nMeta ran the headcount reduction first and hit the bottom of the J-curve with 8,000 fewer people to climb out of it.\n\n## Being Retained Is Not the Same as Being Valued\n\nNow to the part the metrics cannot reach, which is what it does to a person to survive this.\n\nMore than 70,000 people still work at Meta. Many of them spent the first half of 2026 reading press reports about their employer's plan to make itself AI native, being told to install software that recorded their keystrokes and screenshots so agents could learn to do their jobs, watching colleagues disappear in May, and then, in late August, discovering from a Reuters investigation that scenarios had been modelled in which teams like theirs shrank by 60 per cent.\n\nThey were not vindicated. They were retained.\n\nThe distinction matters more than any statistic in this article. To be vindicated is to have your value recognised. To be retained is to have your replacement fail its trial. The reasoning was stated out loud, in internal posts and then in a wire service investigation read by the entire industry: you are still here because the machine broke things, the cleanup was expensive, and the timing was wrong.\n\nThat is a stay of execution, and everyone involved knows it. Zuckerberg made the implication explicit when he said the bets “haven't come to fruition yet”. Nobody at Meta heard that as an abandonment of the thesis. They heard it as a schedule revision.\n\nThink about what this does to the psychological contract. Not the employment contract, which is a document about notice periods, but the unwritten one, the set of mutual expectations that determines whether a person gives an organisation their best judgement or merely their compliance.\n\nDiscretionary effort dies first. The extra hour spent understanding why a system is fragile rather than routing around the fragility. The willingness to raise a problem in a review everybody else wants to wave through. The impulse to mentor somebody junior. All of it is voluntary, all of it is invisible to instrumentation, and all of it depends on believing the organisation intends to keep you.\n\nThen there is the truly perverse incentive, which is the sharpest thing in this whole episode.\n\nTacit knowledge is the asset that saved Meta. The judgement about what not to build, the memory of why a system is shaped as it is, the instinct to refuse. Meta's incident numbers are, on the most sympathetic reading, a measurement of what happens when that knowledge is removed. Any sane organisation reading them would conclude that it must urgently capture and document it.\n\nBut consider the position of the individual asked to document it, inside a company that has already installed keystroke and screenshot capture expressly to train agents to perform white-collar work, and has already modelled cutting teams by 60 per cent. In that context, writing down your judgement is indistinguishable from writing your own replacement's training data. The rational response is to withhold, not out of malice but out of a perfectly reasonable reading of the situation.\n\nSo Meta now needs the thing it has systematically disincentivised anybody from providing. That is not a culture problem a town hall can fix. It is a structural trap of the company's own construction, and precisely the organisational learning failure McClure and Gerdau describe when they argue that AI failure typically presents as technology failure while actually being coordination failure.\n\n## The Ones Who Were Right and Lost Anyway\n\nThen there are the roughly 8,000 people who went in May.\n\nThey were cut on the strength of a thesis: that AI agents could absorb a large fraction of the daily work of the organisation. That thesis has now been falsified, at least for this deployment at this moment, by the company's own internal data, conceded in the company's own all-hands meeting, and the second wave built on it was cancelled.\n\nThey do not get their jobs back.\n\nThis asymmetry ought to sit at the centre of any honest account of Project OT. The evidence arrived after the cost had been paid, and it was paid by people who had no say in generating it. Zuckerberg did not bear the risk of his thesis being wrong. The engineers and designers and programme managers told in April that 10 per cent of the workforce would go bore it. Their mortgages bore it. So did their visa statuses, for anybody on an employment-linked visa, for whom a layoff is not a career interruption but a countdown to leaving the country.\n\nSome are litigating. In July 2026, 26 current and former Meta employees filed suit in federal court in Oakland, California, alleging the company used a constellation of internal artificial intelligence systems to help select employees for the May reduction in force, ranking staff on activity and productivity data including keystrokes, emails, documents, browser history and AI token usage. The plaintiffs allege a disparate impact on employees who had been on protected medical, parental or family leave, since metrics of that kind cannot, by design, be accumulated by somebody who is not at work. Meta has said the claims “lack merit and are not based on facts”, and that “workforce management and organizational decisions were and are made by people, not AI”. Judge William Orrick denied a temporary restraining order on 17 July, finding no irreparable harm while acknowledging “serious questions going to the merits”. At a preliminary injunction hearing on 24 August, he said the plaintiffs were unlikely to win one on the current record, and that their claims appeared better suited to arbitration, where they could pursue their request for an independent audit of Meta's layoff decisions.\n\nWhatever the courts decide, the allegation illuminates something structural. The same measurement apparatus that failed to capture what the humans contributed was allegedly used to decide which humans to keep. Motion counted. Absence of motion counted against you. And the agents, whose entire behavioural repertoire is motion, would have scored magnificently.\n\n## Meta Is Not Alone, Which Is the Problem\n\nThe reversal genre now has a small canon.\n\nIn February 2024, Klarna announced that an OpenAI-powered assistant was doing the work of 700 customer service agents, handling 2.3 million conversations in its first month. In May 2025 its chief executive, Sebastian Siemiatkowski, told Bloomberg that Klarna had cut too far, that lower cost had produced lower quality, and that the company would rebuild human support. “Really, investing in the quality of human support is the way of the future for us,” he said, having spent a year saying close to the opposite.\n\nSalesforce went the other way and stayed there. In September 2025, Marc Benioff said its support workforce had gone from around 9,000 to around 5,000 as AI agents took over roughly half of interactions, with hundreds of affected staff redeployed. Whether that holds is an open question, but it is a genuine counter-example and should not be waved away. Some of these deployments do work.\n\nWhat distinguishes Meta's case is scale and documentation. Klarna's reversal was inferred from customer complaints. Meta's was forced by a quantified collapse in the relationship between output and outcome, recorded in its own systems, at a company with more engineering telemetry than almost any organisation in history. If Meta cannot make the numbers work with that instrumentation, the confident restructuring memos currently circulating in firms with a fraction of its measurement capability deserve far more scepticism than they are getting.\n\n## The Sequencing Was the Crime\n\nSo what went wrong?\n\nNot that the agents were bad. They were, on the evidence, extremely productive in the narrow sense of generating change and merely useless in the broad sense of generating value. That gap will close, possibly quickly. Nothing in the Meta data supports the claim that agentic software development has hit a permanent ceiling, and anybody using this story to argue that AI is fake is reading it lazily. Four months of internal telemetry at one company in 2026, confounded by a simultaneous mass layoff, tells you about a deployment strategy. It does not tell you about 2028.\n\nThe failure was sequencing. Meta removed the people first and measured the consequences second. It made an irreversible decision on the basis of a reversible belief. The layoffs were permanent and the thesis was provisional, and the company got those two properties the wrong way round.\n\nNothing prevented the opposite order. Run the agents at scale for two quarters. Watch the incident rate. Watch the ratio between changes made and features delivered. Establish, with the measurement apparatus you already own, whether the productivity is real or merely legible. Then decide how many people you need. That costs a few months of payroll. The sequence Meta chose cost 8,000 jobs, a nineteen-point collapse in employee sentiment, a federal lawsuit, a 40 per cent rise in major incidents, and a permanent contribution to the industry's growing suspicion that leadership is guessing.\n\nCompanies choose the wrong order not out of stupidity but because the wrong order is legible to markets and the right order is not. Announcing a 10 per cent headcount reduction alongside capital expenditure guidance of 125 to 145 billion dollars is a coherent story about discipline that any analyst can model. Announcing a two-quarter measurement exercise is not a story at all. The sequencing error is a communications strategy that metastasised into an operating decision.\n\nWhich returns us to the ratio.\n\nTwo hundred and twenty per cent against 36 per cent is not a story about machines failing. It is a story about an organisation discovering, at enormous human cost, that it had never known the difference between the work it counted and the work that mattered. The agents did not introduce that confusion. They exposed it, at volume, the way a stress test exposes a flaw that was always present in the drawings.\n\nThe uncomfortable ending is that Meta has not solved this. It has postponed it. The thesis was not withdrawn, only rescheduled, and the measurement problem that made it look plausible in January is exactly as unsolved in September. Nobody at Meta has published a better definition of value. Nobody has replaced the commit count with something that captures refusal, or judgement, or the memory of why a system is shaped the way it is.\n\nSo the next time the agents get good enough, the same dashboard will be consulted, and it will still be measuring motion.\n\nThe people who were kept know this. That is the part the ratio does not capture.\n\n## Sources and References\n\n1. Katie Paul, Reuters, “Mark Zuckerberg had a bold plan to replace Meta staff with AI. Here's how it imploded”, 26 August 2026. [https://www.reuters.com/investigations/mark-zuckerberg-had-bold-plan-replace-meta-staff-with-ai-heres-how-it-imploded-2026-08-26/](https://www.reuters.com/investigations/mark-zuckerberg-had-bold-plan-replace-meta-staff-with-ai-heres-how-it-imploded-2026-08-26/)\n2. Katie Paul and Courtney Rozen, Reuters, “Exclusive: Meta's Zuckerberg says AI agent tech progressing slower than expected”, 2 July 2026. [https://finance.yahoo.com/technology/ai/articles/exclusive-zuckerberg-says-ai-agent-201123441.html](https://finance.yahoo.com/technology/ai/articles/exclusive-zuckerberg-says-ai-agent-201123441.html)\n3. Evan Schuman, Computerworld, “Meta's plans to replace workers with AI fell flat, report says”, 26 August 2026. [https://www.computerworld.com/article/4214479/metas-plans-to-replace-workers-with-ai-fell-flat-report-says.html](https://www.computerworld.com/article/4214479/metas-plans-to-replace-workers-with-ai-fell-flat-report-says.html)\n4. Will Shanklin, Engadget, “Meta reportedly abandoned an AI-focused restructuring plan that would have laid off thousands”, 26 August 2026. [https://www.engadget.com/2244816/meta-reportedly-abandoned-an-ai-focused-restructuring-plan/](https://www.engadget.com/2244816/meta-reportedly-abandoned-an-ai-focused-restructuring-plan/)\n5. Skye Jacobs, TechSpot, “Zuckerberg's bold plan to replace Meta staff with AI has imploded”, 27 August 2026. [https://www.techspot.com/news/113631-zuckerberg-bold-plan-replace-meta-staff-ai-imploded.html](https://www.techspot.com/news/113631-zuckerberg-bold-plan-replace-meta-staff-ai-imploded.html)\n6. The Next Web, “Meta planned to cut some teams by 60% and replace the work with AI”, 31 August 2026. [https://thenextweb.com/news/meta-project-ot-ai-native-layoffs-cancelled](https://thenextweb.com/news/meta-project-ot-ai-native-layoffs-cancelled)\n7. Michael Polanyi, “The Tacit Dimension”, Doubleday, 1966; University of Chicago Press reissue, 2009. [https://press.uchicago.edu/ucp/books/book/chicago/T/bo6035368.html](https://press.uchicago.edu/ucp/books/book/chicago/T/bo6035368.html)\n8. International Business Times UK, “Meta Was Hours Away From Another Massive AI Layoff Wave, Then Zuckerberg Pulled the Plug”, August 2026. [https://www.ibtimes.co.uk/meta-ai-restructuring-zuckerberg-cancels-layoffs-1816629](https://www.ibtimes.co.uk/meta-ai-restructuring-zuckerberg-cancels-layoffs-1816629)\n9. Marco Quiroz-Gutierrez, Fortune, “Meta laid off 10% of its workforce as Mark Zuckerberg warns that in the AI race 'success isn't a given'”, 21 May 2026. [https://fortune.com/2026/05/21/meta-10-percent-workforce-layoffs-ai-tech-success-is-not-a-given-8-thousand-employees-mark-zuckerberg/](https://fortune.com/2026/05/21/meta-10-percent-workforce-layoffs-ai-tech-success-is-not-a-given-8-thousand-employees-mark-zuckerberg/)\n10. Eva Roytburg, Fortune, “Meta will start tracking employees' screens and keystrokes to train AI tools”, 21 April 2026. [https://fortune.com/2026/04/21/meta-will-start-tracking-employees-screens-and-keystrokes-to-train-ai/](https://fortune.com/2026/04/21/meta-will-start-tracking-employees-screens-and-keystrokes-to-train-ai/)\n11. CBS News / Associated Press, “26 Meta workers sue over alleged AI-aided layoffs targeting employees on medical or family leave”, 15 July 2026. [https://www.cbsnews.com/news/26-meta-workers-sue-ai-aided-layoffs-medical-family-leave/](https://www.cbsnews.com/news/26-meta-workers-sue-ai-aided-layoffs-medical-family-leave/)\n12. Margaret Attridge, Courthouse News Service, “No immediate relief for Meta workers claiming AI fired them”, 17 July 2026. [https://www.courthousenews.com/no-immediate-relief-for-meta-workers-claiming-ai-fired-them/](https://www.courthousenews.com/no-immediate-relief-for-meta-workers-claiming-ai-fired-them/)\n13. MLex, “US judge skeptical of Meta workers' AI layoff claims, points to arbitration”, 24 August 2026. [https://www.mlex.com/mlex/artificial-intelligence/articles/2517219](https://www.mlex.com/mlex/artificial-intelligence/articles/2517219)\n14. Sabry E. Farrag, “The Productivity-Reliability Paradox: Specification-Driven Governance for AI-Augmented Software Development”, arXiv:2605.01160, 1 May 2026. [https://arxiv.org/abs/2605.01160](https://arxiv.org/abs/2605.01160)\n15. Won Ik Cho, Seong-hun Kim and Geunhye Kim, “Position: Adopting AI in Practice Does Not Guarantee the Productivity Boost”, arXiv:2605.24688, 23 May 2026. [https://arxiv.org/abs/2605.24688](https://arxiv.org/abs/2605.24688)\n16. Jeanne McClure and Gregg Gerdau, “Why AI Readiness Is an Organizational Learning Problem, Not a Technology Purchase”, arXiv:2604.16369, 22 March 2026, revised 21 April 2026. [https://arxiv.org/abs/2604.16369](https://arxiv.org/abs/2604.16369)\n17. DORA, Google Cloud, “State of AI-assisted Software Development 2025”, 29 September 2025. [https://dora.dev/dora-report-2025/](https://dora.dev/dora-report-2025/)\n18. METR, “We are Changing our Developer Productivity Experiment Design”, 24 February 2026. [https://metr.org/blog/2026-02-24-uplift-update/](https://metr.org/blog/2026-02-24-uplift-update/)\n19. Nathen Harvey and DORA team, Google Cloud, “The ROI of AI-Assisted Software Development (2026.01)”, 11 May 2026, reported by InfoQ, “New DORA Report Claims Strong Engineering Foundations Drive AI Return on Investment”. [https://www.infoq.com/news/2026/05/dora-roi-ai-assisted-dev-report/](https://www.infoq.com/news/2026/05/dora-roi-ai-assisted-dev-report/)\n20. METR, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity”, 10 July 2025. [https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/)\n21. GitClear, “The Maintainability Gap: 2026 AI Code Quality Research”, January 2026. [https://www.gitclear.com/the_ai_code_quality_maintainability_gap](https://www.gitclear.com/the_ai_code_quality_maintainability_gap)\n22. Sherin Shibu, Entrepreneur, “Klarna Is Hiring Customer Service Agents After AI Couldn't Cut It on Calls, According to the Company's CEO”, 9 May 2025. [https://www.entrepreneur.com/business-news/klarna-ceo-reverses-course-by-hiring-more-humans-not-ai/491396](https://www.entrepreneur.com/business-news/klarna-ceo-reverses-course-by-hiring-more-humans-not-ai/491396)\n23. Fortune, “Salesforce CEO Marc Benioff says his company has cut 4,000 customer service jobs as AI steps in: 'I need less heads'“, 2 September 2025. [https://fortune.com/2025/09/02/salesforce-ceo-billionaire-marc-benioff-ai-agents-jobs-layoffs-customer-service-sales/](https://fortune.com/2025/09/02/salesforce-ceo-billionaire-marc-benioff-ai-agents-jobs-layoffs-customer-service-sales/)\n24. Marilyn Strathern, “'Improving Ratings': Audit in the British University System”, European Review, volume 5, 1997, pages 305 to 321. [https://gwern.net/doc/statistics/decision/1997-strathern.pdf](https://gwern.net/doc/statistics/decision/1997-strathern.pdf)\n25. Meta, “Update on Meta's Year of Efficiency”, 14 March 2023. [https://about.fb.com/news/2023/03/mark-zuckerberg-meta-year-of-efficiency/](https://about.fb.com/news/2023/03/mark-zuckerberg-meta-year-of-efficiency/)\n\n**Tim Green**\n*UK-based Systems Theorist & Independent Technology Writer*\n\nTim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at [smarterarticles.co.uk](https://smarterarticles.co.uk), challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.\n\nHis writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.\n\n**ORCID:** [0009-0002-0156-9795](https://orcid.org/0009-0002-0156-9795)\n**Email:** [tim@smarterarticles.co.uk](mailto:tim@smarterarticles.co.uk)\n\nListen to the free weekly [SmarterArticles Podcast](https://www.smarterarticles.fm)", "url": "https://wpnews.pro/news/220-per-cent-more-code-why-meta-stopped-replacing-staff-with-ai", "canonical_source": "https://smarterarticles.co.uk/220-per-cent-more-code-why-meta-stopped-replacing-staff-with-ai?pk_campaign=rss-feed", "published_at": "2026-09-29 01:00:30+00:00", "updated_at": "2026-09-29 01:19:11.944696+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-agents", "ai-safety"], "entities": ["Meta", "Mark Zuckerberg", "Reuters", "Katie Paul", "Courtney Rozen", "Project OT"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/220-per-cent-more-code-why-meta-stopped-replacing-staff-with-ai", "markdown": "https://wpnews.pro/news/220-per-cent-more-code-why-meta-stopped-replacing-staff-with-ai.md", "text": "https://wpnews.pro/news/220-per-cent-more-code-why-meta-stopped-replacing-staff-with-ai.txt", "jsonld": "https://wpnews.pro/news/220-per-cent-more-code-why-meta-stopped-replacing-staff-with-ai.jsonld"}}