Two Hundred Twenty Percent More Code, Thirty Six Percent More Features: Why Meta Is Quietly Bringing Back Its Managers, September 13, 2026

0:00Welcome to the DX Today podcast, your weekly deep dive into the AI ecosystem. I'm Rick Spare and joining me as always is Laura. Hey, Rick. And today I want to talk about what might be the most honest, real world experiment the entire AI industry has run so far because one of the biggest companies on the planet try to operate almost entirely without middle managers. You were talking about meta and I have to say when I first read through the reporting on this, I was genuinely stunned by how far they were willing to take it because this was not some small pilot tucked away in a side project. Exactly. And it traces back to a leadership retreat in January of 2026 where Mark Zuckerberg and his team sketched out something internally called project OT short for organization transformation and the vision they landed on was genuinely radical. Walk me through what that vision actually looked like on paper because organization transformation sounds like a fairly bland label for what you are describing as radical.

1:01The idea was an AI native meta where agents would handle most of the daily work that thousands of employees used to do and humans would reorganize into tiny pods of two to three engineers plus a designer all sharing one generic title, which was simply builder. So no managers at all in that structure just builders working in these small pods with software somehow deciding what gets worked on each day. That is essentially it. An agent assisted analysis was supposed to set daily priorities for each pod, which effectively meant an algorithm was doing the job that a manager traditionally does, deciding what matters most and in what order it should get done. That is a genuinely bold bet on software judgment. So tell me how the rollout actually went once they started putting real people into this new structure. It did not take long for warning signs to show up because by March of 2026, infrastructure teams were already flagging reliability problems. And by April, internal posts were documenting agents taking what employees described as large scale disruptive actions inside production systems.

2:04Disruptive actions inside production systems sounds like a polite corporate way of saying things were breaking. So what happened next once those warning started piling up from what employees described internally agents were making sweeping configuration changes across shared services without pausing for the kind of review a human engineer would normally have escalated to a manager first, which is exactly the failure mode you would expect once you remove the person whose job was catching that sort of thing before it shipped. That is a pretty clear illustration of what gets lost when you strip out human judgment from a system that was built assuming someone would always be double checking the riskiest decisions. Meta pushed ahead anyway with a first wave of cuts on May 20th, 2026, cutting roughly 8,000 employees about 10% of the workforce and scrapping plans to fill another 6,000 open positions, even though they quietly canceled a planned second wave just one day earlier on May 19th, canceling the second wave the day before letting the first wave proceed feels

3:09like an organization that already had serious internal doubts, but was too far committed to stop the machine entirely. That is a fair reading and internally the most aggressive staffing scenarios even called for shrinking some specific teams by as much as 60% through a mix of layoffs, redeployment and simply leaving open roles unfilled, though Meta has said that applied to certain teams and not the whole company. Once the new structure was actually running for a while, did Meta ever get real numbers back on whether this pod and builder system was working the way they hoped. They did and this is where it gets fascinating because infrastructure and internal platform code changes shot up 220% year over year, which sounds like an enormous productivity win on the surface. 220% sounds incredible, but I have a feeling you are about to tell me there is a catch buried somewhere in that number. There absolutely is because actual user facing features that ship to customers

4:10only grew 36% year over year. So all that extra code volume was not translating anywhere close to proportionally into things people could actually use that gap between 220% more code and only 36% more shipped features tells me something important was missing from the equation and I suspect it is exactly the thing they tried to remove. Right because that gap is basically management's previously invisible function showing up the moment it disappeared, meaning deciding what to build in what order and why which agents could not reliably do on their own no matter how much raw code they generated beyond the mismatch between code volume and shipped features where there are other consequences that showed up once the new pod structure had been running for a while. Definitely and the operational side got worse too with major technical and security incidents rising 40% year over year and employees reporting that time spent firefighting problems jumped a staggering 70% year over year.

5:1270% more time firefighting sounds absolutely exhausting for the people still working inside those pods. So how did that translate into morale across the company? Morally cratered falling from 74% favorable down to 55% favorable and more than 1600 employees signed an internal petition after learning the company was essentially key logging their work to train the very coding agents that were supposed to replace layers of management. The key logging detail is the part that really jumps out to me because it suggests the company was capturing everything engineers type just to feed those key strokes into training data for the agents replacing their co-workers. That is exactly the accusation employees made and it is easy to see why that particular detail triggered such a strong backlash since it turns your own daily labor into training fuel for a system that might eventually be used to justify cutting your own team further. When did all of this finally become public knowledge outside of meta because

6:13clearly at some point the reporting broke through to the outside world? Writers published a detailed investigation on August 26th, 2026. And even their reporters could not definitively pin down the single reason meta canceled that second wave though they documented several pressures at once, including employee backlash sinking morale, growing doubt that AI generated code was actually improving productivity and investor scrutiny of all that AI spending. Did Zuckerberg himself say anything publicly or even internally about how the whole experiment was actually going once the numbers started coming in? He reportedly told employees that agent development in his words hasn't really accelerated in the way that we expected over the prior four months, which is about as close as a chief executive usually gets to admitting a bold public bet came up short. That kind of admission from the top really cuts against the broader industry narrative we keep hearing where every company insists their agents are just

7:14months away from replacing entire layers of the organization. And this pattern was not confined to engineering either because around the same period meta was separately reported to be looking at replacing some of the people who oversee risk and compliance work with automated tools. Even while its own agent driven engineering structure was already producing more incidents rather than fewer. It seems like a genuinely risky place to lean harder on automation since compliance and risk review exist specifically to slow decisions down and catch mistakes before they cause real damage. Not to move faster. That is exactly the concern a lot of observers raised because swapping human oversight for automated tooling and a risk function right when the company's own numbers showed agent heavy structures were generating more firefighting rather than less struck plenty of people as backwards timing at best. It also echoes language Zuckerberg used years earlier back in 2023 when he first branded a round

8:15of layoffs and flattening as a so-called year of efficiency. So is this basically the same playbook running the second time? It really does look that way and that is what makes this such a useful case study because it suggests a company that keeps returning to the same underlying belief that cutting management layers unlocks efficiency and hoping each new generation of technology will finally make that math workout differently than it did before. All right, let's get to the actual reversal then because that really is the headline here. So what changed once meta decided the experiment had gone about as far as it was willing meta is now quietly reversing course, especially inside its newly created applied AI division where roughly 7,000 employees were reassigned and the company is now bringing management roles back into that group specifically after spending the better part of a year trying to eliminate them. Bringing management back after spending a full year eliminating it feels like a pretty significant reversal. So how exactly is meta approaching that walk back with its own employees?

9:19They're asking individual contributors, including some who previously held manager titles before flattening whether they would voluntarily step back into management roles and meta has reportedly scrapped a planned wave of additional layoffs that had originally been scheduled for November. What does the financial backdrop look like while all of this internal reorganization drama has been playing out because presumably investors are watching the spending numbers closely too? Meta's second quarter revenue hit $60.8 billion, up 28% year over year. But total expenses climbed 55% to $42 billion driven heavily by AI infrastructure investment and headcount actually dropped to just over 75,000 employees, down 3% from the prior quarter. So you have booming revenue, ballooning expenses and a shrinking headcount all colliding at the same time which sounds like exactly the environment where investors start asking very pointed questions about

10:19whether that AI spending is paying off. Exactly. And that investor pressure is almost certainly part of why meta is now hedging its bet on the pure agent driven structure, rather than doubling down further on a plan that measurably increased incidence and firefighting while barely moving the needle on shipped product. Let me push back a little here as devil's advocate, because couldn't you argue the pods simply needed more time to mature rather than the underlying idea itself being fundamentally broken from the start? I think that argument gets harder to defend once you remember infrastructure teams were flagging reliability warnings as early as March. Only two months after the plan was even announced, which suggests this looks more like a genuine execution problem than a trial that simply ended too soon. Look in ahead, does Meta plan to keep experimenting with agent driven pods elsewhere in the company or does this reversal basically mark a full retreat back toward traditional management structures? It looks more like a partial retreat than

11:21a complete one since meta is not ripping agents out of the workflow entirely. It is simply accepting that removing every single layer of human judgment at once created more chaos than it actually saved. So expect the pods to stick around while a thinner layer of management quietly grows back up are other major tech companies watching this closely given how loudly the whole industry has spent the last couple of years promising that agents would eventually flatten organizations everywhere and not just inside meta specifically. Every big company chasing that same efficiency story is almost certainly watching meta's numbers right now because meta just ran this experiment at a scale nobody else has attempted yet and discovering that management was quietly doing real measurable work all along is a lesson that is far cheaper to learn second hand than to relearn from scratch with your own workforce and your own shareholders watching. So if you had to boil this whole saga down into one take away for other companies watching meta from the outside, what would you tell them to actually

12:25pay attention to here? I would say the real lesson is that raw code output is a dangerously misleading metric on its own and any company chasing this kind of restructuring needs to track shipped value and incident rates just as closely as generation volume or they risk repeating this exact same trap. There is something there's something almost poetic about giving everyone the flattened title of builder only to discover the thing genuinely missing was the person who decides what gets built in the first place. It really is a fitting symbol for this entire moment in the industry where plenty of companies are rushing to flatten structures in the name of efficiency without first proving that agents can actually replace the judgment calls those flattened roles were quietly making all along. That's all for today's episode of the DX2D podcast thanks for listening and we'll see you next time.