Two of your Developers now direct AI agents, and each agent works from a skill file: the written instructions for one type of work. One Developer found her agents writing code without proper error handling and fixed her skill file that afternoon. Nobody checked whether the other Developer’s files carried the same gap. Meanwhile, someone above your teams has proposed an AI oversight group to review the AI agent skill files across the agency. Before that group is built, look at what the Scrum Master accountability you already hold says about those files.
In Brief
- Each Developer who directs AI agents owns the content of the skill files those agents work from, and revises them when output falls short.
- The Scrum Master stewards the files as a system, keeping them consistent with each other and with the Definition of Done.
- A gap that recurs across Developers’ skill files is an impediment the Scrum Master acts on in the current Sprint.
- Skill file problems the team can’t fix travel to whoever holds operating model effectiveness across the teams, often an existing director.
Developers own their skill files
When AI enters a Scrum Team, each Developer can delegate routine work to agents, against the team’s Definition of Done and under its skill files (Hodgson & West, 2026b). The Developer keeps the novel, high-judgment work. They inspect their agents’ combined output against the Definition of Done and revise the skill file when the output falls short. That sits squarely inside a line you already coach to: the Scrum Guide makes the Developers accountable for instilling quality by adhering to a Definition of Done (Schwaber & Sutherland, 2020). The Developer’s accountability extends to their agents’ output, and nobody gains a new accountability.
The Developer’s authority stops at what Hodgson and West (2026b) call the irreversibility boundary: the point where a decision can’t be undone without disproportionate cost. Integrating a component into the team’s Increment is the Developer’s call. Releasing a breaking change to production is the Product Owner’s.
Each Developer sees only their own agents. In the example Hodgson and West (2026b) give, a Sprint Retrospective finds agents producing code without adequate error handling, and the Developer whose agents did it updates the skill file. From the next Sprint, every agent working to that file handles errors. The same gap can still sit, worded differently, in another Developer’s files and produce the same failure there.
Scrum Masters steward the skill file system
Paper 5 of the Empiricism at Machine Speed series separates stewardship from ownership: each Developer owns the content of the skill files their agents use (Hodgson & West, 2026b). The Scrum Master is accountable for the system those files form: whether they stay consistent across Developers and each conforms to the Definition of Done and the rules of Scrum. You’re also accountable for whether the revision process works, and for when a change made for one Developer’s agents should become a team-wide standard.
The Scrum Master doesn’t govern any Developer’s agents or take over their files. You govern how the standards written into the instructions change, and the agents keep their own ways of breaking down and sequencing their work (Hodgson & West, 2026b).
The Scrum Guide makes you accountable for the Scrum Team’s effectiveness and asks you to help the team focus on high-value Increments that meet the Definition of Done (Schwaber & Sutherland, 2020). Once agents do much of the work, the skill files are where the Definition of Done reaches it, so your accountability reaches the files too.
Each Developer owns what their skill files say. The Scrum Master is accountable for whether the files still work together.
Paper 5 presents this model as a hypothesis with no operational evidence behind it. No published case yet describes a Scrum Team whose Developers each govern their own agents. The one team the series studied had a single human Product Owner governing eight agents (Hodgson & West, 2026b). Treat the model as something your teams test Sprint by Sprint, with the evidence in front of you.
Scrum Masters act on drift this Sprint
Skill files drift even when nobody edits them. In Evolve, Hodgson (2026) calls this instruction drift: an agent follows stale rules with full confidence and produces compliant output that no longer has customer value. A change of model version, including one the provider makes without your agency asking for it, changes how every instruction is read while the text stays the same.
In my Zombie Scrum talk, I said that changing people’s behaviour is not about changing people, but changing the context which they are in. Agents make that literal: their skill files are their context, and the agents’ output only improves when someone changes the files.
One signal Hodgson (2026) names is a skill file with no revision history across several Sprints while the product’s context keeps moving. Another is files that have drifted apart from each other and from the Definition of Done, where no single agent is failing and the divergence shows only when you look at the output together. Recording the model version in each skill file, and checking recent work against the Definition of Done whenever it changes, turns an invisible shift into something the team can inspect (Hodgson & West, 2026a).
When you find a gap that recurs across Developers’ files, raise it in the Sprint you find it. The Scrum Master serves the team by causing the removal of impediments to its progress (Schwaber & Sutherland, 2020). A gap that produces the same failure in two Developers’ files is an impediment. The Sprint Retrospective, where the Scrum Team inspects its tools and its Definition of Done, is where Paper 5 puts the regular check that the skill files still work together (Hodgson & West, 2026b).
Some skill file problems need escalating
Some skill file problems can’t be fixed inside one Scrum Team, such as a standard several teams share or a model change rolled out across the agency. The product operating model in Evolve gives these their own path (Hodgson, 2026). Value questions travel from the Product Owner to a Product Manager, who holds value direction for the whole value stream. Operating model questions travel from the Scrum Master to a Product COO, the person accountable for operating model effectiveness across the teams in that value stream. That accountability is often attached to an existing role, such as a public service director (Hodgson, 2026).
A skill file problem that travels the wrong path gets the wrong fix. When it travels the value path, it surfaces as a performance problem, and the corrective action addresses the wrong layer (Hodgson, 2026). When it travels no path at all, nothing reaches the people funding the work.
Hodgson (2026) describes an Australian federal agency whose product operating model cut decision latency from several months to one day. Within nine months, traceability from strategic objectives to the teams’ work rose from 10% to 92%. When the coaching investment was withdrawn, conditions degraded within six months. The cadence held while the accountability structure didn’t, and nothing was checking whether the conditions that made it work were still in place. Skill files fail the same quiet way, while every Scrum event still runs on time.
What this means for Scrum Masters: When your Developers direct AI agents, each of them owns what their skill files say, and you steward the system those files form. Your part is keeping the files consistent with each other and with the Definition of Done, and acting on a recurring gap in the Sprint you find it. Questions the team can’t settle go to whoever holds operating model effectiveness above it. Share this with your community of practice before anyone builds a separate AI oversight group. A Scrum Team that stewards its own skill files fixes drift where it starts, and the people above it get the questions only they can answer.
References
- Hodgson, M. (2026). Evolve: The operating model AI demands. Zen Ex Machina. https://zenexmachina.com/evolve/
- Hodgson, M., & West, D. (2026a). Empiricism at machine speed, Part 4: Failure modes for agentic systems [Whitepaper]. Scrum.org. https://www.scrum.org/resources/empiricism-machine-speed-5-part-series
- Hodgson, M., & West, D. (2026b). Empiricism at machine speed, Part 5: The agentic nexus [Whitepaper]. Scrum.org. https://www.scrum.org/resources/empiricism-machine-speed-5-part-series
- Schwaber, K., & Sutherland, J. (2020). The Scrum Guide. Scrum.org. https://scrumguides.org/scrum-guide.html