160 M C K I N S EY Q UA RT E R LY where cross-functional teams already under stand modern software engineering. - Invest in knowledge graphs. Knowledge graphs are essential because they unify all the informa tion inputs—code repositories, documents, and more—into a single structured network that shows how concepts, facts, and assets are connected. - Learn to break work into agent-ready tasks. Humans need to develop the skill of decompos ing larger features into small, well-scoped tasks with clear inputs, outputs, and acceptance cri teria. This is what allows multiagent workflows to run safely. Without discrete, agent-ready work items, agents either stall or drift. - - Master spec-driven development and context engineering. Teams need to get very good at defining clear specifications—what the system should do, how it should behave, and how it will be tested. AI can generate code, but only when its instructions are crisp, structured, and complete. Equally important is giving agents the right context—architecture diagrams, data models, APIs, constraints, and business rules—so the AI can make correct decisions. Good AI output comes from good context, not clever wording. Strengthen human judgment and review skills. Humans become the editors in chief of the fac tory. They must review proposed updates, catch architectural drift, assess whether the agent’s work matches intent, and decide when to tighten guardrails or adjust tests. This combination of product judgment, architectural understanding, and quality review remains fully human. - Revisit performance expectations. Human– agent productivity changes how teams operate. LATAM found one of the biggest challenges to adopting agentic AI was redeploying people into additional tasks as agents freed up time. Some companies reduce team size; others raise the bar on what should be delivered in a quarter. Either way, operating models must shift. Monitor token consumption closely. In a world where teams can spin up agents, which then create additional prompts or spawn subagents, token consumption can grow exponentially and lead to significant cost overruns (tokens are essentially processing units for LLMs). To counter this issue, build up your financial oper ations (FinOps) management to track and direct agent activity. - Running an AI agent factory is not about swap ping humans for automation; it is about creating the conditions where humans and AI agents can work together at high speed and with quality. The human capabilities that sit on top—decomposing work, exercising judgment, tuning the system, and managing cost—are what will turn an agent factory from an experiment into a durable advantage. - What would happen if progress in software devel opment productivity moved from the current frontier of two-times improvement to the new frontier of 20-times improvement? How would this change the world of business? - We know the road there might be a bit bumpy— any one of us can point to issues companies have been having even making modest productivity improvements with AI agents. But this world is coming, and senior executives should be thinking through scenarios and implications. » Charlotte Relyea is a senior partner in McKinsey’s New York office, and Martin Harrysson is a senior partner in the Bay Area office. » Excerpted with permission from the publisher, Wiley, from Rewired: How Leading Companies Win with Technology and AI , by Eric Lamarre, Kate Smaje, and Rob Levin, with Alex Singla and Alexander Sukharevsky. Copyright © 2026 by McKinsey & Company. All rights reserved. This book is available wherever books and ebooks are sold. AI can generate code, but only when its instructions are crisp, structured, and complete.