The scrum master's hidden job, and the AI teammate built to take it

A scrum master is the person on an agile software team accountable for how the team works. The job description includes coaching, clearing obstacles and keeping delivery honest. The hours go elsewhere: reviewing requirements, checking the board, chasing updates, flagging priorities and moving information from where it lives to where a decision needs it.

None of this is hard, but it crowds out the actual job. The 2020 Scrum Guide is specific about what that job is. The scrum master is accountable for establishing Scrum and for the team’s effectiveness, a “true leader who serves”: coaching self-management, causing the removal of impediments, helping the product owner with goals and the backlog, leading the organization’s adoption of Scrum. 

Every item on that list is judgment work. Status reconciliation, ticket cleanup and relaying blockers from chat to board appear nowhere in it. That work is accretion. Tool sprawl created it, and the scrum master absorbed it because someone had to. The cost hides because the work looks like productivity. A scrum master assembling the standup summary appears fully occupied. The question is whether the team hired a coach or an information aggregator.

The usual fixes each treat a symptom. Dashboards add visibility, but visibility was never the bottleneck. A human still opens them, decides what matters and acts in another system. Rule-based automation handles narrow, well-specified conditions, and sprint reality is not one of them. 

“Flag tickets that look blocked” is not a rule. It is a judgment weighing comment tone, elapsed time and dependencies. More headcount puts expensive, trained judgment back onto work that never needed it. All three assume a human must operate the work, or a rigid rule must encode it. None of these hold true anymore.

The Scrum community has weighed in. The 2025 Scrum Guide Expansion Pack, co-authored by Scrum co-creator Jeff Sutherland, invites teams to create agents as AI team members, and it is equally plain about the limits: the scrum master must be human, and people stay accountable for all outcomes. Agents on the team, judgment with the people.

So, we built a teammate to take that job. Known as Kai, it has its own name on the roster.

A human operates a tool, so the tool creates work by demanding attention. A team delegates to a member, so the member absorbs work by owning an outcome. A digital team member is an AI agent with its own identity in the team’s systems. This digital colleague carries the coordination work autonomously and pulls a human in only for decisions that need one.

Identity is the part most teams skip. In Gravitee’s State of AI Agent Security 2026 survey of more than 900 executives and practitioners, 88% of organizations reported confirmed or suspected agent security incidents in the past year. Yet, only 22% treat agents as independent, identity-bearing entities. An agent on borrowed credentials stays a tool. The human owns every action and every consequence.


Kai gets what a new hire gets. Its own entry in the identity provider, with a named human sponsor. Its own email, calendar, Slack handle and seat on the Jira board. Onboarded like a hire, offboarded the same way, every token revocable in seconds. 

Because it acts under its own account, every board update is attributable to Kai or to a person. It even answers standup voice notes in a synthesized voice of its own. Every action runs through the Model Context Protocol, so the tools underneath are swappable. With that standing, the team could assign Kai work, and the question shifted from what we automated to what we delegated.

Identity is necessary but not sufficient. Three more decisions did the rest.

The agent operates on two rhythms and remembers. Standups, ceremony scheduling, blocker sweeps and risk checks run on a cadence: fifteen workflows today. Questions get answered all day: ticket status, workload, sprint trajectory. An agent that runs only rituals is a newsletter. One that only answers questions is a search bar. And a flag informed by three sprints of history reads differently from one triggered by a threshold. Context changes the meaning. 

Work reaches it the way work reaches a person. No control panel, no dispatcher. Coordination lives where the 2020 guide puts it: in the board, ceremonies and roles. Work moves through board states, and whoever owns the next state picks it up—human or agent. 

When Kai hands a ticket to a developer agent, the handoff is a board transition with checkable exit criteria, the Definition of Done, and machine honored. No central orchestrator means no single point of failure, and the next agent needs no reconfiguration.

Security got more design effort than intelligence did. An agent with standing identity in your systems is a real attack surface, and pretending otherwise is how these projects die. Kai authenticates with short-lived credentials, and every tool call passes a gateway that validates the token first. A guardrail layer wraps every model call and action. The team can tighten it but cannot switch it off. 

Outside text, an email or a ticket body, can never trigger a tool call on its own, and any human can pause any agent from chat. Nothing consequential happens alone: before Kai assigns a ticket or changes a status, it waits for human approval. The agent aggregates and proposes. The human judges and approves.

Attention moves first. Standup prep, ceremony scheduling, twice-daily blocker sweeps, workload checks and scope-drift comparisons happen automatically. Questions that once required opening three tools get answered in seconds because the agent already has the context.

Mapped against the guide’s accountabilities, the division of labor is clean. Coaching the team in self-management and leading the adoption of Scrum stay fully human. Three are shared: Kai preps sprint data, the human facilitates the events. Kai surfaces impediments around the clock, the human removes them. Kai handles backlog hygiene, the human advises on technique. 

Keeping the artifacts visible is the one job Kai does almost end to end, with the human still accountable for the result, exactly where the Expansion Pack draws the line. Judgment stays human. Accretion moves.

Capacity moves second. The agent compresses each team down to the handful of decisions that need a person, so a scrum master is no longer capped by the boards they can personally watch. Ours can coach four or five teams instead of hand-coordinating two, at a fixed monthly cost in the low hundreds of dollars for around ten thousand requests.

There are two honest caveats: None of this eliminates the scrum master. It changes what the hours buy, so a team that measures the role by ticket-edit volume will need a new measure. And a team that rubber-stamps every approval has quietly switched off the safety mechanism. The design assumes an engaged human, and that’s exactly the point.

Nothing that makes Kai work is specific to coordination. The same foundation of identity, memory, approval gates and secure system access supports a roster of persona-based teammates we call CAST (Cognitive Autonomous Software Teammates). Eden, the developer persona, implements tickets tagged agent-eligible, writes the tests and opens the pull request. 

Merging is always a human’s call. A QA persona tests work in a sandbox and files bugs with reproduction steps, built on a different model family so its review does not share the developer’s uncertainties. A product owner persona drafts tickets and synthesizes feedback but escalates every priority and scope call to the human because fully autonomous product ownership is the wrong design.

Humans and agents share one board and its history shows exactly who did what. The scrum master is the proof that the pattern holds, the first hire from a roster you can extend to any role.

Start with a diagnostic, not a rollout. For one week, count the coordination actions your scrum master performs: every status reconciliation, every blocker relayed from chat to board, every ticket touched to keep things honest. Sort them into two piles, work that needed judgment and work that only needed attention. The size of the second pile is the size of the opportunity.

Then, stage the rollout. Let the agent own the scheduled rituals first, where work repeats and risk stays low. Add question-answering once the team trusts the reports. Add approval-gated actions when the approval flow feels natural. The roster grows the same way, one persona at a time, each earning its place before the next joins.

The bigger shift: teams are moving from software that they operate to colleagues they delegate to. Early movers will stop asking what they can automate and start asking what they would hand to a capable, always-available teammate, and what judgment they would keep. That second question is the entire design brief. The technology to answer it exists now.