The Agentic SDLC Operating Model: Team Topologies, Talent, and Change Management

At one global systemically important bank, McKinsey researchers describe a London office where the day crew of three engineers arrives at 8 a.m. to find nearly a hundred AI agent teams have just finished an overnight shift, having refined a cross-border payment system, tested failure paths, and shipped updates at a pace no human team could match. The engineers don’t start a two-week sprint cycle. They start a sprint review, because with agents working continuously, that review now happens every morning. The result, per McKinsey: ten times the speed at half the cost.

That’s not a story about faster code generation. It’s a story about a different-shaped team, running on a different cadence, with a different definition of what the humans in the room are actually there to do. Most conversations about agentic AI in software engineering focus on the tools. The structural questions, how teams are sized, who fills which role, and how governance gets built into the workflow, are the ones that actually determine whether an organization gets the G-SIB bank’s result or another stalled pilot.

Most current SDLC team structures, cross-functional squads, two-pizza-sized teams, stream-aligned and platform teams, were designed around human coordination limits: how many people can share context effectively, how many handoffs a team can absorb before velocity drops. Those constraints don’t disappear with agentic AI, but the unit they’re being applied to changes.

McKinsey’s research on the agentic organization describes teams of two to five people already supervising an “agent factory” of 50 to 100 specialized agents running an entire end-to-end process, whether that’s onboarding a customer, launching a product, or closing the books. Applied to software delivery, that’s not a team topology with an AI tool added to it. It’s a materially different-sized atomic unit, with humans positioned “above the loop” to steer outcomes rather than execute every step, selectively stepping back into the loop only where human judgment or contact specifically matters.

The organizational chart itself is described as shifting shape. Rather than the hierarchical delegation a traditional org chart represents, McKinsey’s research describes agentic organizations pivoting toward “work charts,” structures based on exchanging tasks and outcomes across a flat network of agentic teams rather than reporting lines. For an SDLC specifically, that shows up exactly the way the G-SIB bank example does: the daily rhythm of the team reorganizes around what the agents produced overnight, not around a fixed ceremony calendar built for a slower cycle.

McKinsey’s agentic organization research identifies three roles emerging as people work alongside agents: M-shaped supervisors, broad generalists fluent in AI who orchestrate agents and the hybrid workforce across domains; T-shaped experts, deep specialists who reimagine workflows, handle exceptions, and safeguard quality; and AI-augmented frontline workers, who spend less time operating systems directly and more time on the judgment calls that still require a person.

Inside a technology organization specifically, McKinsey’s research on designing the AI-first technology workforce describes the central hiring question shifting from how to scale engineering capacity to how to allocate human talent where judgment and decision-making still matter. Demand is rising for senior engineers, architects, product managers, and designers who can define standards and orchestrate development work across internal teams, vendors, and agents, while the business case for hiring large numbers of junior developers weakens as agentic AI absorbs the execution work that used to justify those roles.

That doesn’t mean the talent pool for these roles is limited to people who were already senior engineers. McKinsey’s agentic organization research notes early evidence that employees without technical backgrounds, citing a French literature graduate on one of its own project teams, have proven able to learn agentic workflow management as quickly as trained engineers. The skill that matters most in these new roles is closer to end-to-end problem-solving and system-level judgment than to any specific programming background, which broadens who can credibly move into an M-shaped or T-shaped role even as it narrows the traditional junior-engineer entry path.

McKinsey’s research draws a direct parallel to how DevSecOps embedded automated security and compliance checks directly into the delivery pipeline rather than treating them as a separate audit function. Agentic organizations are described as extending that same logic to agent oversight itself: critic agents that challenge outputs, guardrail agents that enforce policy, and compliance agents that monitor regulatory requirements, all built into the workflow rather than bolted on afterward.

Human accountability doesn’t disappear under this model, but its shape changes. Instead of line-by-line code or output review, the human role becomes defining policy, monitoring outliers, and adjusting how much oversight a given workflow needs. McKinsey’s research is candid about the risk in getting this wrong in either direction: too little oversight creates real exposure, but too much oversight caps the productivity gain, since the scale of agentic adoption is ultimately limited by how much oversight capacity humans can actually provide. For an SDLC, that means code review, security review, and compliance sign-off need to be redesigned as agent-supervision functions with clearly defined escalation paths, not preserved as unchanged manual gates sitting in front of agent-generated output.

McKinsey’s research frames the shift required as three deliberate moves, not a rollout schedule. Organizations evolve linearly by default while the underlying technology is improving exponentially, so leaders have to make a bold operating-model change rather than a series of incremental ones. The transformation can’t be delegated to a technology leader the way a software deployment would be, since it requires the business to envision the future operating model and work backward, not wait for IT to hand over a finished platform. And it has to be treated as an opportunity to engage employees continuously, with real investment in upskilling, incentives, communication, and performance management, rather than a threat to be managed quietly.

The scale of the gap most organizations are starting from is larger than it might seem. McKinsey’s research puts 89 percent of organizations still operating on industrial-era hierarchical models today, with 9 percent running agile or product-and-platform models from the digital era, and only 1 percent operating as a decentralized network. Most engineering organizations attempting an agentic SDLC transition are further back on that curve than their leadership typically assumes, which is exactly why DORA’s research describes AI as an amplifier: it strengthens organizations with strong existing engineering practices and just as reliably exposes the ones without them. Change management here isn’t a soft add-on to the technical rollout. It’s the variable that determines whether the same AI investment produces the G-SIB bank’s result or an amplified version of existing dysfunction.

A few concrete moves follow from where the evidence actually points. Size human teams around supervision ratios rather than headcount-reduction targets, using a pattern like McKinsey’s two-to-five-people-per-50-to-100-agents as a reference point, then adjusting to actual outcome ownership rather than a fixed percentage cut. Redesign the ritual layer, standups, sprint reviews, and release cadences, around the pace agents can actually produce work at, rather than preserving a cadence built for a slower, human-only cycle. Build governance into the workflow structure itself, with defined agent roles for critique, guardrails, and compliance, rather than layering a manual review step in front of agent output after the fact. Hire and reskill for orchestration and exception-handling judgment specifically, and expect the pool of people who can credibly fill these roles to be broader than “senior engineers only.” And treat the whole effort as an operating-model change owned by business and engineering leadership together, with sustained investment in upskilling and communication, not a project handed to IT to roll out quietly.