“AI-native project management” is a phrase that can mean almost anything and therefore usually means very little. I am more interested in a smaller question: what may an agent do to delivery work without quietly becoming the process owner?
In Tracker, an agent can help with bounded actions around a Story. It can inspect context, prepare a change, reorder work within the rules or report what happened. The action has a known shape, a target and a state transition. That is less magical than letting an agent roam around the project, which is probably why it is more useful.
The boundary matters because project data is not harmless decoration. Reordering work changes attention. Moving a Story changes expectations. Marking something finished changes what people believe. Those actions need explicit permissions and an audit trail, not a cheerful sentence from a bot saying it took care of things.
The current work around Tracker is pushing that shape further through bounded delivery commands and stricter reorder and MCP schemas. The details are deliberately unexciting: validate the input, enforce the allowed transition, record the result and fail clearly when the requested action does not fit.
A human still owns acceptance. An agent may prepare, check, summarize and carry out an allowed action. It does not get to declare the outcome valuable just because the command returned 200. “The script ran” and “the work is accepted” are different facts. Systems should keep those facts apart.
This also gives teams a better answer to the security question. Do not ask whether an agent is trusted in the abstract. Ask which action it may take, on which object, under which state, with which evidence and who can undo or reject it. Smaller questions produce controls you can actually review.
The goal is not an autonomous digital scrum team. That sounds like a staffing plan written by a toaster. The goal is to remove repetitive delivery work while keeping intent, scope and accountability visible to the people doing the work.
If Tracker earns its place, an agent should make the team faster without making the team guess what changed. That is a much better product promise than pretending the board can now run itself.