# Moving away from Agile: What's Next Martin Harrysson & Natasha Maniar, McKinsey & Company | AI Engineer CODE 2025 | 21:55 Source: https://www.youtube.com/watch?v=SZStlIhyTCY Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/moving-away-from-agile-whats-next Published: 2025-12-12 Tags: engineering-culture, product-strategy, team-adoption ## TL;DR - AI tools produce limited enterprise-wide gains when companies leave team structures, roles, review processes, and planning methods unchanged. - AI-native organizations use smaller pods, continuous planning, spec-driven development, and workflows that assign different roles to humans and agents. - Scaling AI requires hands-on upskilling, incentives, measurement of business and engineering outcomes, and many coordinated changes across the organization. ## Summary Martin Harrysson and Natasha Maniar argue that adding AI tools to existing Agile teams leaves much of the available value unrealized. Individual developers may save hours, while companies report only modest overall improvement because work allocation, code review, collaboration, team design, and role definitions have not changed. They describe AI-native workflows with continuous planning, spec-driven development, smaller pods, consolidated product-building roles, and different operating models for legacy modernization and new feature work. Examples from enterprise clients include agent-assisted story assignment, prototypes for clearer acceptance criteria, workflow-based squads, and real-time customer feedback. The speakers also stress that organization-wide adoption depends on hands-on coaching, incentives, certifications, and measurement. Their framework tracks inputs, delivery speed, developer experience, code quality, resilience, and economic outcomes. They recommend starting now, testing the model that fits each organization, and setting a larger ambition than adding another coding assistant. ## Key ideas ### AI productivity stalls when companies keep the old operating model [00:13](https://www.youtube.com/watch?v=SZStlIhyTCY&t=13s) Martin Harrysson says AI has created dramatic improvements for individual tasks, such as work that once took days now taking minutes. Yet a survey of about 300 mostly enterprise companies found average overall productivity improvements of only 5, 10, or 15 percent. Teams are generating more code, but many still review it manually. AI-generated code can also add technical debt and complexity. Work allocation has become harder because agents perform very differently across tasks, and people have different levels of experience with them. The result is a gap between what AI can do in isolated cases and what companies gain across software development. ### AI changes the bottlenecks in collaboration and review [04:04](https://www.youtube.com/watch?v=SZStlIhyTCY&t=244s) As development speeds up in some areas, collaboration among people and teams often stays the same. Stories written in prose with detailed acceptance criteria can produce code that does not match the intended result, leaving manual review as the main control. Harrysson describes this as automating some work while creating more review work. Managers also face a difficult allocation problem because some tasks work very well with agents and others do not. The speakers connect these bottlenecks to the persistence of eight-to-ten-person teams, two-week sprints, and other parts of the Agile operating model built around older constraints. ### Different software work needs different human-agent operating models [06:41](https://www.youtube.com/watch?v=SZStlIhyTCY&t=401s) Natasha Maniar says rewiring the product development life cycle cannot follow one universal design. Legacy modernization needs broad codebase context and clearly defined outputs, so a factory of agents can use a human-provided initial specification and a final human review with little intervention. Greenfield and brownfield feature work benefits from variation and non-deterministic outputs, so an iterative loop may work better. In that model, agents act as co-creators that offer options and support faster feedback. The operating model should therefore follow how humans and agents collaborate on the type of work being done. ### AI-native companies change workflows and team structure together [08:43](https://www.youtube.com/watch?v=SZStlIhyTCY&t=523s) In their survey of 300 enterprises, the speakers found that top performers were seven times more likely to have AI-native workflows that covered more than four software life-cycle use cases. They were six times more likely to have AI-native roles, including smaller pods with different skills. These organizations moved from quarterly planning toward continuous planning and from story-driven work toward specs that product managers iterate on with agents. They also described one-pizza pods of three to five people, with product builders managing agents and working across the full stack. Product managers are beginning to create prototypes directly in code instead of iterating only on long product requirements documents. ### Enterprise teams can test smaller pods and agent-assisted delivery [10:57](https://www.youtube.com/watch?v=SZStlIhyTCY&t=657s) In a study with a leading international bank, team leads assigned sprint stories with agents using team velocity and delivery history. Teams then co-created prototypes and refined acceptance criteria around security and observability, which reduced downstream rework. Squads were reorganized by workflow, with separate focus areas for small bug fixes and greenfield development. Agents checked possible cross-repository effects in the background. Product managers also used real-time customer feedback to reprioritize features rather than waiting for data scientist input. The reported results included more than a 60-times increase in agent consumption and a 51 percent increase in code mergers, alongside faster delivery tied to business priorities. ### Roles must be rewritten because AI changes what engineers and product managers do [12:11](https://www.youtube.com/watch?v=SZStlIhyTCY&t=731s) The speakers say engineers are moving from mainly executing and writing code toward orchestrating agents and deciding how to divide work between people and tools. They report that about 70 percent of surveyed companies had not changed software roles, even though they expected employees to work differently. Experiments with smaller pods consolidated tasks previously split across separate roles. This allowed more pods to operate with the same number of people while maintaining or improving code quality. The speakers also describe product managers creating code prototypes and taking a more direct role in shaping software outputs. ### Scaling AI depends on coordinated change management [15:39](https://www.youtube.com/watch?v=SZStlIhyTCY&t=939s) For organizations with hundreds of teams, software tools alone do not scale the change. Harrysson describes change management as getting many small things right at the same time, including communication, incentives, and upskilling. One technology company saw AI-tool usage fall off or remain ineffective after rollout, even as more users were added. The reset included clearer expectations for developers and product managers, hands-on upskilling with bring-your-own-code sessions, available coaches, and support during the first few sprints. Code labs and certifications also helped another client change day-to-day behavior. ### Measurement should connect AI investment to engineering and business outcomes [18:43](https://www.youtube.com/watch?v=SZStlIhyTCY&t=1123s) The speakers recommend measuring more than tool adoption. Their framework starts with inputs such as spending on AI tools and the time devoted to training and change management. It then tracks adoption, velocity, capacity, developer experience, code security and quality, and resilience. Mean time to resolve priority bugs is one proxy for resilience. At the business level, measures can include time to revenue, the price difference for higher-quality features, customer expansion, and cost reduction per pod. They say bottom-performing enterprises often failed to measure speed, and only 10 percent measured productivity. The right proxies will change as the tools develop. ## Notable quotes - "There is a bit of a disconnect between this big potential around AI and the reality." (02:54) - "So you've automated some things but we've generated more manual reviews." (06:13) - "This is a human change, and it takes some time, and it's a big change." (21:08) - "Building a robust measurement system that prioritizes outcomes and not just adoption is important not just to monitor progress, but also pinpoint issues and course-correct quickly." (18:20) ## Tools & references mentioned - McKinsey & Company - Software X - Agile - Kanban - Carnegie Mellon - Cursor ## Who should watch - Engineering leaders whose teams use coding assistants but see only modest delivery gains will find concrete causes to investigate. - Product and engineering managers planning smaller teams or agent-based workflows can compare the proposed models with their current setup. - Enterprise transformation leaders need a measurement and adoption plan that covers training, incentives, developer experience, quality, and business results. ## Related talks - [AI Leadership](https://aietalks.com/talks/ai-leadership) (Alex Lieberman, Morning Brew and 10X & Kath Korevec, Google Labs & Katelyn Lesse, Anthropic & Michele Catasta, Replit & Lisa Orr, Zapier & Steve Yegge, Sourcegraph and AMP & Gene Kim, IT Revolution & Bill Chen & Brian Fioca, OpenAI & Martin Harrysson & Natasha Maniar, McKinsey & Yegor Denisov-Blanch, Stanford & Itamar Friedman, Qodo & Olive Song, MiniMax & Asaf Bord, Northwestern Mutual & Lei Zhang, Bloomberg & Samir Mody, The Browser Company & Max Kanat-Alexander, Capital One & Arman Hezarkhani, 10X & Justin Reock, DX & Dan Shipper, Every & Mel Lutzky, Graphite, 8:16:05) - [Leadership in AI Assisted Engineering](https://aietalks.com/talks/leadership-in-ai-assisted-engineering) (Justin Reock, DX, 18:11) - [Unlocking AI-Powered DevOps Within Your Organization](https://aietalks.com/talks/unlocking-ai-powered-devops-within-your-organization) (Jon Peck, GitHub, 22:13) - [How AI Is Changing Software Engineering](https://aietalks.com/talks/how-ai-is-changing-software-engineering) (Gergely Orosz, The Pragmatic Engineer, 26:42) - [Everything We Knew About Software Has Changed](https://aietalks.com/talks/everything-we-knew-about-software-has-changed) (Theo Browne, @t3dotgg, 16:02)