Cut the Middle, Keep The Mess

Cutting the middle layer looks like an easy win. Trim headcount, flatten the org chart, hand the work to AI agents, bank the savings. The real story is messier, and it starts with what actually breaks when the middle disappears.

On the latest episode of Deep Geeks, WEKA's Val Bercovici sits down with host Serena Huang to explain why the job everyone assumed AI would make obsolete might actually be the job AI needs most.

Middle management's bad rap

Middle management has been an easy target for a decade. Every earnings call has a version of the same line: we eliminated waste by cutting that inefficient layer. It sounds disciplined. It plays well with investors. And now, as companies rush to replace that same layer with AI agents, the math isn't landing the way anyone expected.

Val's read on it: look at what the best-performing AI agents actually do. They organize in swarms. They take an ambitious goal, split it into subtasks, delegate the work, monitor quality, and synthesize everything back into a finished result. That's not a technical skill. That's a classic MBA middle-management skill, and it turns out to be one of the hardest things to get an AI system to do well.

Watch: Val explains why agent swarms need an org chart

Why "replace the manager" backfired

Serena speaks about how some companies didn't just get the capability gap wrong. They got the cost wrong too, and cost is the part of this story most leadership teams aren't measuring correctly.

AI agents don't work like chat. A single ambitious project run through an agent swarm can rack up tens of thousands of dollars in token costs, and unmanaged, that number climbs fast. Someone still has to decide which task gets a premium model and which gets a bargain-bin one, someone still has to catch a project running in the wrong direction before it burns through the budget, and someone still has to translate the finished output into something the business actually needed. Remove the person doing that job, and the agents don't get cheaper. They get more expensive and less accurate at the same time.

The real bottleneck isn't the org chart

Val's argument runs deeper than staffing. It isn't just that companies need people managing AI. It's that the entire industry underestimated what it actually costs to run these systems well, and the root cause isn't leadership decisions. It's memory.

Every AI agent swarm depends on massive amounts of high-bandwidth memory to hold context, and that memory is in short supply industry-wide. Val makes the case that this shortage, not model quality, is the real constraint on how big and how smart these agent systems can get. He uses an analogy involving a stove, a prep station, and a kitchen that can't keep up no matter how expensive the equipment is. Once you hear it, you'll understand why a badly managed agent swarm and a badly resourced kitchen fail for the exact same reason.

Watch: The kitchen analogy that explains the memory wall

Cheaper tokens, bigger bills

Falling token prices don't mean falling bills. Val breaks down the math, and it explains why some companies are already looking at five-figure daily AI costs, even though the per-token price kept dropping the whole time. Without someone managing that tradeoff deliberately, the savings never show up. The bill just gets bigger.

Watch the full episode

If you've spent the last few years being told your management skills are becoming obsolete, this episode makes the opposite case. The skills that make a good manager, setting clear goals, delegating intelligently, catching problems early, and measuring what actually matters, are turning out to be exactly what makes an AI system worth its cost.

Deep Geeks Ep. 3 goes past the layoff headlines and into what's actually driving AI cost and capability right now: memory scarcity, tokenomics, and the management skills making an unexpected comeback.

If someone on your team is questioning whether their job survives the next wave of AI, share this with them.

Listen to this episode on Spotify or Apple or watch on YouTube. Make sure to subscribe to Deep Geeks to get notified of future episodes that cover AI infrastructure.