Orchestrating Agents Was Supposed to Be the New Job. The Agents Are Catching Up.

Designing AI for human inclusion, ten slides

Save it, share it, or send it to whoever just got told to “learn to orchestrate the agents.”

Swipe to preview · 10 slidesDownload the PDF

The management layer being cut

The reskilling advice and the org chart are moving in opposite directions. Manager headcount at large public companies fell 6.1% between May 2022 and May 2025, with executive roles down 4.6% over the same stretch, according to Live Data Technologies data reported by Fortune. Estée Lauder and Match Group each cut roughly a fifth of their managers. Amazon eliminated around 14,000 corporate roles in October 2025, with CEO Andy Jassy explaining the logic directly: managers “want to put their fingerprint on everything,” he said, and the result is “people being in the pre-meeting, for the pre-meeting, for the pre-meeting, for the decision meeting.” Cloudflare cut about a fifth of its workforce in May 2026 while posting record revenue, and CEO Matthew Prince named exactly who the cuts targeted: “measurers,” his term for middle management, finance, legal, internal audit, and revenue recognition roles. A Korn Ferry survey of 15,000 professionals globally found 41% now say their own company has trimmed management layers in the past year.

None of these companies are framing the cuts as “we automated the manager’s job.” They’re framing it as removing a layer of coordination and approval that AI tooling made unnecessary. That’s a more precise and more uncomfortable claim than a simple automation story, because it means the job wasn’t replaced by a system doing the same tasks. It was removed because the reason for the layer to exist, keeping information moving and decisions checked between people who couldn’t otherwise see the whole picture, stopped requiring a person to do it.

The capability curve behind that decision keeps compounding

METR’s March 2025 research gives the clearest measurement of why this isn’t a one-off restructuring wave. Tracking task length by how long a skilled human takes to complete it, METR found the length of task frontier AI agents can complete autonomously at 50% reliability has doubled roughly every seven months, consistently, for six years. At the time of publication, models were near-perfect on tasks that take a person under four minutes, and under 10% reliable on tasks that take a person more than four hours. Orchestration, in the sense the reskilling advice means it, sits inside that longer-horizon band: deciding what to delegate, sequencing dependent steps, catching a bad output before it compounds. It’s exactly the kind of task the doubling curve is climbing toward, not a category exempt from it.

McKinsey’s own estimate of what generative AI changed captures the same shift in a single comparison. In McKinsey Global Institute’s pre-generative-AI 2017 estimate, “managing and developing talent” carried a 16% automation potential. By its 2023 reassessment, done after generative AI’s language and reasoning capabilities were factored in, that figure had risen to 49%. The activity most closely associated with what a manager does, not the spreadsheet work around it, is the one where the estimate moved the most.

Reskilling into orchestration assumes orchestration stays a human role

The advice to become the person who directs agents rather than the person who does the task isn’t wrong as a near-term move. It’s incomplete as a long-term one, because it treats “orchestrator” as a stable destination rather than another point on the same curve everything else has been sliding down. A market-entry analyst who spends two years learning to sequence agent workflows well is building a skill with the same shelf life as the task it was built to escape, if the pattern holds. The mid-level manager told to become a reviewer of AI output rather than a producer of it is being handed the review layer, which is precisely the layer Amazon and Cloudflare just cut because a person checking a person’s work, or a person checking an agent’s work, is the coordination cost both companies decided they no longer needed.

That’s the part the reskilling framing tends to skip. It answers “what should I learn” without answering “for how long does learning it matter,” and the honest answer right now is that nobody knows, including the people selling the framing with confidence. Beautiful.ai’s 2026 survey of 3,000 US managers found 72% believe their own employees fear AI will make them less valuable at work, up eight points from the year before, and 70% believe their employees fear AI will eventually cost them their job outright, up twelve points. Managers aren’t reporting confidence in the orchestration pivot. They’re reporting that the anxiety it was supposed to resolve is getting worse, not better, among the exact population being told to make that pivot.

Wisdom work was never only about task execution

The task-replacement framing that shapes most AI product design treats a manager’s job as a bundle of tasks: write the report, run the analysis, draft the review, schedule the sync. Automate enough of the bundle and the job looks solved. What that framing leaves out is the part of the job that was never a task in the first place: being the person whose judgment a decision is attached to, who can be asked why, who bears the consequence if the call turns out wrong, and whose presence in the room is itself what makes the outcome legitimate to the people affected by it. A capable agent can produce the analysis. It cannot, in any way that currently means anything to the people relying on the decision, be the party accountable for having made it.

That distinction is not sentimental, and it is not a permanent moat either. It is a design decision about where a system draws the line between output and accountability, and right now most AI systems are built to blur that line rather than mark it, because marking it is slower and less impressive in a demo. A system designed for task replacement optimises for getting a human out of the loop as fast as it can. A system designed for human inclusion has to decide, on purpose, which decisions keep a specific person load-bearing even when the system is capable of producing the same output alone, and build the workflow so that person’s judgment is structurally required.

Coexistence needs specific design choices now

There is no clean end state to point to here, and it would be dishonest to pretend otherwise. Nobody currently building or deploying these systems has proof of where the automation frontier stops, or whether it stops at all before it reaches the accountability layer itself. What’s within reach is narrower and more useful: the systems being built and bought right now can be designed to keep a specific human decision load-bearing at a specific point, or they can be designed to route around that point the first time it’s inconvenient. That choice is being made today, by product teams and by the companies buying their tools, whether or not anyone frames it as a choice.

A small number of concrete practices follow from taking that seriously. Build the checkpoint into the workflow, not into a policy document, because a checkpoint someone can skip under deadline pressure isn’t a checkpoint. Name the person accountable for a given class of decision before the system ships. Measure how often the human checkpoint changes the system’s output, not just whether it exists: a rubber-stamp review that never overturns anything is a compliance artifact. Reject the framing that reskilling into “orchestrator” is a permanent answer, and treat every new capability jump as a fresh occasion to ask which decisions still deserve a structurally required human in them. None of this stops the capability curve. It decides what the curve is used for while it climbs.

The uncomfortable truth in the brief that prompted this post is the right one to sit with: we can train, upskill, and repurpose people faster than most organisations currently manage, and it may still not be enough, because the frontier keeps moving to wherever the retraining just arrived. That’s not a reason to stop trying. It’s the reason the harder, less comfortable work has to happen at the design table, deciding which decisions a system is built to need a person for.

If you’re deciding where a human decision needs to stay structurally required in an AI system you’re building or buying, that’s exactly the kind of design call my consulting work helps sequence.