The Agentic Organization: What McKinsey's Next Operating-Model Shift Means for How You Build Software

From Farms to Factories to Agents

Every few generations, the dominant way work gets organized resets. McKinsey's 2025 thesis on the agentic organization frames this as a sequence of operating models, and the history is a useful anchor before we talk about what changes.

In the agricultural era before the 1800s, operating models were small teams of craftspeople and farmers — 80 to 90 percent of the global population worked in agriculture. The industrial era moved people into factories and reorganized them into functional hierarchies; products were designed for mass replication, with major upgrades every three to ten years. By the 1970s, 39 percent of US workers were in the industrial sector and only 4 percent in agriculture. Efficient scaling became the source of competitive advantage, and lean management became a strategic tool.

The digital era, from the 1990s onward, broke the industrial maxims. Monolithic systems gave way to modular products and platforms shipped monthly, weekly, or — at the extreme — continuously. Amazon famously releases code every 11.6 seconds. Speed demanded agile operating models: small cross-functional teams of software engineers, designers, and product managers. Today roughly 5.8 percent of the US population works in tech, 19.3 percent in industry, and 1.6 percent in agriculture.

McKinsey's argument is that we are now at the front edge of a fourth reset. The previous eras revolutionized physical work. This one revolutionizes knowledge work — and the organizing unit becomes a network of humans and AI agents working together.

The Capability Curve That Makes This Plausible

It's easy to dismiss "agents will run your company" as vendor theater. What makes McKinsey's version worth reading is that it hangs the whole thesis on a measurable trend rather than vibes.

Citing METR's research on AI task length, the report notes that the duration of tasks an AI can reliably complete on its own has been doubling roughly every seven months since 2019 — and every four months since 2024 — reaching about two hours of equivalent work as of writing. If that curve holds, AI systems could complete around four days of work without supervision by 2027.

Read that as a trajectory, not a prophecy. The honest framing is conditional: capability growth depends on continued model progress, on regulation, and on whether the curve bends. But even the conservative reading reorders priorities. An agent that reliably owns a two-hour task is already past the "autocomplete" stage and into the "junior teammate who needs review" stage. The interesting question for an engineering leader stops being "can AI write code" and becomes "what supervision structure do I need so that work done by agents is safe to ship?"

That's the shift that matters. The bottleneck moves from generating output to verifying, integrating, and owning it — which is a human-and-process problem, not a model problem.

The Five Pillars of an Agentic Organization

McKinsey structures the redesign around five pillars of the enterprise. They're worth knowing because they map cleanly onto questions any technology leader is already asking.

Business model: with agents able to deliver services at near-zero marginal cost, the products and markets you can credibly serve widen. The report's example is a bank that extends from mortgages into furnishing, renovations, and energy upgrades — because the agentic capacity to serve those journeys suddenly exists.

Operating model: the org chart stops being a hierarchy of people and becomes a decentralized network of outcome-focused agentic teams — humans and agents grouped around a result, not a function.

Governance: when agents act, accountability, audit trails, escalation paths, and guardrails become first-class design problems rather than afterthoughts.

Workforce, people, and culture: roles shift from doing the task to supervising, orchestrating, and judging the work of agents — which raises, not lowers, the premium on senior judgment.

Technology and data: the substrate. Without clean, accessible, well-governed data and the platform plumbing to let agents act safely, the other four pillars are slideware.

What an Agentic Workflow Actually Looks Like

The report's clearest illustration is a bank. A customer wants to buy a house. A personal AI concierge activates a chain of agentic workflows: a real-estate agent suggests properties, an underwriting agent tailors mortgage offers to the customer's financial profile, compliance agents check the deal against bank policy, a contracting agent finalizes the agreement, and a fulfillment agent funds the loan. Crucially, the whole chain is overseen by an agentic team of human supervisors, mortgage experts, and AI-empowered frontline staff.

Translate that into software delivery and it stops being abstract. Picture a feature moving through a pipeline where a planning agent decomposes the ticket, an implementation agent writes the code, a test-generation agent builds coverage, a review agent flags risks, and a deployment agent ships behind a flag — with a senior engineer owning the outcome, arbitrating the judgment calls, and signing off on anything that touches money, data, or security.

The point of both examples is the same: the work is decomposed into agent-executable steps, but a human team owns the result. The bank doesn't fire its mortgage experts; it repositions them as supervisors of a system that now moves far faster. That repositioning — not the automation itself — is the hard part.

Virtual Agents, Physical Agents, and a Spectrum of Ambition

McKinsey is careful to describe a spectrum rather than a binary. Virtual agents are being deployed along increasing levels of complexity: simple tools that augment existing activities, then end-to-end workflow automation, then full "AI-first" agentic systems where the workflow is designed around agents from the start.

In parallel, physical AI is emerging — "bodies" for AI in the form of smart devices, drones, self-driving vehicles, and early humanoid robots that let AI act on the physical world. For most software organizations the virtual end of the spectrum is where the next two years live, but the framing matters: the same orchestration-and-supervision discipline applies whether the agent is shipping code or moving a pallet.

The practical takeaway is to locate yourself honestly on that spectrum. Most teams are at "augmenting existing activities" and quietly pretending they're further along. There's nothing wrong with starting at augmentation — but you should know which rung you're on, because the governance and team design needed for an AI-first workflow are an order of magnitude more demanding than for a coding assistant.

What It Means for How You Staff and Build Software

Here's where the thesis lands for engineering leaders, and where our own experience building teams diverges usefully from the consulting framing.

First, the agentic organization is brutal on the headcount-arbitrage model. If a senior engineer plus a set of agents can own outcomes that used to require a squad of mid-level developers, then "we have 40 people on this account" stops being a selling point and starts being a liability. The value migrates to small, senior-dense teams that can supervise agents, make architectural judgment calls, and take accountability for results. This is the same structural pressure we've written about in agentic engineering and the collapse of the junior developer pipeline — McKinsey's framework is the org-design view of the same trend.

Second, governance is not a compliance checkbox you bolt on later — it's the thing that determines whether agentic velocity is an asset or a time bomb. Agents that ship fast without audit trails, escalation paths, and a human owner produce exactly the kind of structural debt that takes years to repay. The teams that benefit from agents are the ones that build the supervision layer first.

Third, the human roles that survive and thrive are the senior ones. "Supervise, orchestrate, and judge" is not entry-level work. It requires people who have shipped enough production software to know what good looks like, what's about to break, and when the agent is confidently wrong. That is precisely the profile of engineer we staff — and it's why we've never built our model around volume.

The agentic organization, in other words, doesn't make great engineers less valuable. It makes them the whole game.