Jared Moon, Rory Walsh and Vito Di Leo, with Adam Thelwall, all McKinsey partners, describe in May 2026 how agentic AI is transforming software delivery, which they present as the leading edge of a broader shift in the delivery model. The article opens with a scene: at 9 a.m., a product owner finds that overnight a feature has moved from structured requirements to tested code, with edge cases flagged and architecture dependencies validated. No one worked late; the agents did. According to the authors, several companies are thereby achieving productivity gains of three to five times, with teams reduced by 60%, provided the operating model is redesigned for round-the-clock human-agent collaboration.
First lever: the daily sprint. At night, a factory of agents carries out the structured work; during the day, humans review the outputs, resolve ambiguities, reinforce guardrails and align stakeholders. Their role becomes overseeing the system that produces the artifacts. This presupposes a clear product vision, a standardized technical environment, a uniform requirements-to-code path and stable stakeholders.
this should not begin with a grand, top-down ontology effort
Second lever: eliminating human handoffs. CI/CD mainly automates testing and deployment, roughly 30% of spend; the remaining 70%, from requirements to code, stays manual and interpretive. Codifying specs, non-functional requirements, guardrails and sequence diagrams into machine-readable formats lets the pipeline run end to end in a few hours, with humans stepping in only at defined review points.
Third lever: a knowledge infrastructure. Domain-specific graphes de connaissance serve as agents' memory across the entire software lifecycle, linking customer feedback, architecture decisions, tickets, GitHub activity and incidents. An agent « bibliothécaire » answers in minutes what used to require weeks of interviews with experts, and makes every decision traceable. The authors advise against a top-down ontology: the graph should grow organically around priority domains.
Fourth lever: resizing teams and the portfolio. Teams of eight to twelve people give way to pods of three or four skilled profiles supervising the agents; the illustrative example goes from 200 to 100 person-years and from 24 to 18 months. This requires reskilling, integrating control functions (risk, legal, procurement) via policy as code, and deliberately reallocating the freed-up capacity. The authors recommend starting with the large programs that concentrate spend, and expect human roles to concentrate on architecture, product judgment and systems design.