DORA 2025, the Psychology of Agentic Coding, and Value Stream Management

On This Podcast

In this conversation, we examine how value stream mapping surfaces hidden friction points across a whole software delivery process, from a bank drawing its pipeline on a glass window to a client whose CI/CD pipeline was buckling under five merges nobody had noticed. Rob explains that DORA's 2025 research found value stream management to be a genuine force multiplier for teams using AI: when a team can see its bottlenecks, generating more code doesn't just add more pull requests to an already congested review queue, it lets teams target the improvements that actually move throughput. Daniel adds a real-world example of a QA team overwhelmed once development sped up through agentic coding, illustrating how speeding up one stage of a pipeline strains every other stage that hasn't kept pace. We also discuss how psychological safety within a team is a precondition for these mapping exercises to work honestly.

In this conversation, we examine the central finding of Rob's psychology masters thesis: developers interviewed described a shift in identity from writing code line by line to conducting AI agents through higher-level system thinking. Rob describes his own exhaustion after a few hours directing an agentic coding tool, comparing it to the mental fatigue of his very first pair-programming session with Daniel nine years earlier. We also discuss how this shift forces metacognition, thinking about how you think, as engineers learn to frame intent and context for an agent rather than typing syntax directly, and how some participants found themselves reading architecture books for the first time in years. Rob and Daniel consider what this means for junior developers, who risk losing the learning by osmosis that came from sitting near experienced engineers, and debate whether this is simply another layer of abstraction, like the historic move from assembly code to compiled languages.

In this conversation, we examine Rob's finding that developers are redefining productivity around impact and total cost of build rather than volume of code generated, with several participants explicitly rejecting lines-of-code as a meaningful measure. We also discuss the danger this creates: as AI raises what feels achievable in a day, personal and organisational baselines for productivity quietly rise too, so that anything less than the new elevated pace starts to feel like underperformance. Rob shares that even researching this subject didn't protect him from a bout of burnout partway through his own studies, and Daniel recalls an engineer on a client team who burned out after running multiple parallel Claude Code sessions and lost the ability to switch off. Both agree the responsibility falls to managers and leaders to notice when people are taking on too much and to build in space for learning rather than demanding full utilisation.