AI Engineer Code 2025

Moving away from Agile: What's Next?

Loading slides…

Read the talk

Moving away from Agile: What's Next?

Faster coding exposes slower coordination, review and planning. Capturing the gains requires changing how teams allocate work, define acceptance criteria and measure outcomes.

From a talk by Martin Harrysson and Natasha Maniar

Before you start: Familiarity with Agile sprints, code review and product-management roles will help; no knowledge of a particular AI tool is required.

When the tools change, the work changes

Almost twenty years before this talk, Martin Harrysson joined a technology company as an entry-level software engineer while it was switching to Agile. Kanban boards, stand-ups and new ceremonies changed how the company developed software. Adopting a different method was a substantial organizational change, not simply a new tool on each developer’s machine.

Timeline pairs mainframes and PCs with waterfall, web and client-server with agile, cloud and APIs with product and platform development, and AI coding assistants with AI-native development.Open full source frame
Technology breakthroughs and software development methods, from waterfall to AI-native development.

AI raises a similar operating-model question: what needs to change around the developer when execution becomes much faster? Harrysson and Natasha Maniar approach it through their work with enterprise software organizations. The progression is from individual productivity to team performance, then from successful teams to value across an entire organization.

1:181:30
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

1:18 · section reference included

Why individual speedups fail to scale

The familiar AI success story concerns a bounded task: something that previously took hours or days now takes minutes. Yet the enterprise result can look much less dramatic. Harrysson reports that a survey of about 300 mostly enterprise companies found overall productivity improvements often described as only 5%, 10% or 15%. The gap is between accelerating particular activities and improving the performance of the whole delivery system.

As execution accelerates, other constraints become visible. Collaboration still runs at its previous pace. More generated code enters a review process that remains largely manual. Harrysson also cites Carnegie Mellon research when warning that generated code can amplify technical debt and complexity in some cases. The mechanism is cumulative: faster production creates more material for people to coordinate, understand, review and maintain. Code generation can move the bottleneck without removing it.

Two rising curves show output potential above output reality. Five bottlenecks list collaboration overhead, cognitive limits, manual review and debugging, communication gaps, and increased code complexity.Open full source frame
Process bottlenecks separate output potential from output reality.
2:282:45
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

2:28 · section reference included

Allocation and acceptance criteria become rate limiters

Work allocation becomes harder because AI effectiveness varies along two dimensions. Some tasks respond extremely well to agents; others do not. People also differ in their experience with the tools. An engineering manager therefore cannot treat either tasks or developers as interchangeable units of capacity. Assigning work well requires understanding both the task’s suitability for AI and the person’s ability to direct it.

Review has a related upstream problem. Give an agent a prose story with fuzzy acceptance criteria, and the returned code may differ from what the team intended. If manual review is the only control, automating implementation creates additional human inspection and correction. Better acceptance criteria matter because they constrain the work before the team pays for downstream rework.

The surrounding structure often remains unchanged: eight-to-ten-person teams working in two-week sprints, with practices shaped by the constraints of human implementation. Harrysson describes client experiments over the preceding year with smaller teams, new roles and shorter cycles. The reported performance improvements motivate a broader redesign, but the next step is to choose a collaboration model appropriate to the work itself.

5:025:14
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

5:02 · section reference included

Match human–agent collaboration to the task

Reworking the product development life cycle, or PDLC, does not imply one universal human–agent arrangement. Maniar contrasts two kinds of work whose context requirements and desired outputs lead to different operating models.

WorkAgent contributionHuman involvement
Legacy modernizationA factory of agents works toward defined outputsInitial specification and final review
New features in new or existing productsAgents generate alternatives in an iterative loopOngoing co-creation and feedback

Legacy modernization can require context from the entire codebase while still having a well-defined destination. That combination makes it a candidate for substantial agent execution between human checkpoints. Feature development can benefit from variation: agents offer options, and people use those options to discover what should be built. In that setting, nondeterministic output can support faster feedback rather than merely introduce uncertainty.

7:367:53
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

7:36 · section reference included

Change the workflow and the role together

Maniar reports that surveyed top performers were seven times more likely to have AI-native workflows and six times more likely to have AI-native roles. These are survey associations, not measured causal effects of reorganizing a team. The workflow distinction concerns scaling AI across several parts of the software development life cycle, rather than stopping at an isolated coding or review tool. The role distinction concerns smaller pods with different skills and responsibilities.

Supporting those changes requires continuous, hands-on upskilling, impact measurement and incentives for developers and product managers to adopt AI. Maniar also describes a five-to-sixfold improvement involving time to market and delivery speed, alongside higher quality and more consistent artifacts. The multiplier’s metric and baseline are not defined clearly enough in the talk to use it as an established speedup.

The concrete workflow change begins with planning: move from quarterly planning toward continuous planning, and from story-driven work toward spec-driven development. Product managers iterate specifications with agents instead of repeatedly expanding long product requirements documents. This brings the definition of the work closer to the system that will help implement it.

The corresponding talent model uses three-to-five-person pods. Instead of maintaining separate QA, front-end and back-end responsibilities inside each team, broader product-builder roles combine full-stack fluency, architectural understanding and agent orchestration. Product managers also create prototypes directly in code. Maniar points to Cursor’s internal practices, described in the speakers’ companion article, as an example of these changes operating together.

8:368:57
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

8:36 · section reference included

Inside a bank’s redesigned sprint

For an enterprise already organized around Agile, the transition can begin inside the sprint. In a study with a leading international bank, the team tested changes to the sequencing of work and the division of responsibility between humans and agents. The interventions addressed allocation, rework and collaboration delays together.

  1. Assign work using delivery history. Agents helped team leads allocate sprint stories using team velocity and past delivery data.
  2. Resolve requirements before implementation. Teams co-created multiple prototypes and iterated acceptance criteria with agents, including security and observability requirements. The purpose was to produce more consistent artifacts and reduce repeated corrections during coding.
  3. Organize squads by workflow. Small bug fixes and greenfield development received different squad arrangements, reflecting their different work patterns.
  4. Inspect cross-repository effects in the background. Agents examined potential impacts across repositories to reduce later debugging.
  5. Shorten the customer-feedback path. Product managers observed real-time customer feedback directly and reprioritized features instead of waiting for data-science input. This reduced meeting and coordination delays and helped the backlog advance within the same amount of time.
Sprint-cycle diagram connects refinement, sprint planning, development, sprint review and retrospective, with green intervention markers and callouts for agents, prototypes, squad organization, review and feature prioritization.Open full source frame
AI interventions across an international bank’s sprint cycle.

Maniar reports more than a 60-fold increase in agent consumption and a 51% increase in code merges in the bank study. She also describes faster delivery tied to the bank’s business priorities and improved efficiency. The talk does not specify the consumption unit, measurement window or comparison controls, so the numerical results describe this client intervention rather than a forecast for another bank.

10:4311:04
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

10:43 · section reference included

More pods with the same people

Explicit role design matters because directing agents is different work from executing every implementation step. Engineers increasingly divide work among agents and orchestrate their contributions. Product managers’ responsibilities change as well. Yet Harrysson reports that about 70% of surveyed companies had not changed their role definitions. An expectation of different behavior can therefore coexist with a job description that still describes the old work.

Another client tested smaller pods in place of conventional two-pizza teams. The new roles consolidated responsibilities that had previously belonged to separate specialists. With the same number of people, the organization could form more pods, each expected to perform at approximately the level of its former, larger team. That expectation is central to the capacity argument: the organization gains more independently productive units, rather than merely making each unit smaller.

Harrysson reports maintained—and in some cases improved—code quality, together with faster team output. The case connects role consolidation to additional parallel delivery capacity while keeping quality part of the result.

12:5613:11
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:56 · section reference included

Recovering a stalled rollout

Scaling this approach can mean changing how hundreds of teams and thousands—or tens of thousands—of people work. Harrysson identifies the management of that change as a major difference between organizations seeing modest gains and those seeing much larger improvements. Communication, incentives and upskilling have to reinforce one another. His description of change management is practical: getting twenty, thirty or more small things right together.

At one technology company, tools were rolled out across several parts of the product development life cycle. Initial usage then dropped off or remained ineffective. Adding more users did not change the overall impact. Access to the tools had expanded, but productive habits had not followed.

The recovery began by resetting expectations for what developers and product managers should do differently each day. Training became hands-on: people could bring their own code, and coaches were available to help. The first few sprints received particular attention because that was when new practices needed to become habits. A measurement system made changes and improvements visible.

A separate client used code labs and new certifications to motivate changes in daily practice. Each intervention could appear modest in isolation; their value came from making the desired behavior easier to learn, repeat and recognize across the organization.

15:1715:36
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

15:17 · section reference included

Measure the chain from investment to economic outcomes

Adoption is an intermediate measure, not the final outcome. A measurement system should reveal problems early enough to correct them. Maniar reports that bottom-performing enterprises in the survey were not measuring speed, and only 10% were measuring productivity. Without those measures, an organization cannot tell whether wider tool use is improving delivery.

The proposed measurement chain begins with investment and follows its effects through engineering performance:

LayerWhat to track
InputsAI-tool spending; time and resources for training and change management
Adoption and capabilityBreadth and depth of tool use; upskilling
DeliveryVelocity and capacity
Developer experienceDeveloper NPS; enjoyment or frustration
Engineering healthSecurity, quality and resilience

Velocity and capacity are only part of the picture. Developers may produce more while finding their work more frustrating, or the code may become harder to sustain. Maniar gives mean time to resolve priority bugs as one client’s proxy for resilience, making the connection to operational behavior concrete.

Three rows show investments in tools, training and change management; adoption, upskilling and developer experience; and outcomes of velocity, capacity, security, quality and resiliency.Open full source frame
A measurement framework connects inputs and outputs to enterprise outcomes.

The next layer connects engineering changes to executive priorities. Faster delivery matters economically when it changes the time needed to reach a revenue target. Better features may support a price premium or attract additional customers. Lower human-labor requirements may reduce cost per pod. These are distinct paths to value, so a single coding-productivity measure cannot represent all of them.

Those economic gains can, in turn, fund more greenfield and brownfield development. The framework is a starting point for following that chain, not a fixed set of proxies for every future tool. As agents and workflows evolve, the measures used to represent their effects will need to evolve as well.

18:3118:53
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

18:31 · section reference included

Start the human transition

Predicting the next five years of agent capability is difficult. Maniar’s proposed direction is nevertheless intended to survive improvements in both agent intelligence and human AI fluency: shorter sprints, smaller teams and a larger number of teams. The enduring design question is how to organize people and agents so that faster execution becomes faster, better delivery.

Harrysson closes by urging organizations to start because human change takes time. Choosing a model that fits the organization and setting a bold ambition belong together: the destination should be consequential, while the working arrangement must fit the work. The tools can arrive quickly; changing responsibilities, habits and expectations is the longer transition.

20:3420:52
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

20:34 · section reference included

Resources

Transcript download
Full transcript

Loading transcript…