A project manager opens the tracker on a Monday and reads forty-one merged changes from a team of five — three developers, each running agents over the weekend on work they queued up Friday afternoon. The traceability matrix, the spreadsheet that maps requirements to the work that satisfies them, was last touched Thursday. By the time Monday is reconciled, Tuesday has happened.
Here is the thesis: when agents collapse the build cycle, the slowest artifact in the delivery organization becomes the status report — and the answer is not writing it faster, it is not writing it at all.
The speed change is not anecdotal. Gartner expects 60% of organizations to run smaller engineering teams at scale by 2029, up from 15% now.56 BCG describes delivery setups where as few as three engineers run the factory, with a human accountable at every stage gate.58 McKinsey’s phrase for the professional’s new job is supervising the systems that produce the artifacts, rather than producing the artifacts.57
Every one of those sentences is about engineering. Almost nobody is writing about what the same shift does to the people accountable for the plan.
The PM’s whole toolkit assumes human pace
A weekly status cycle assumes a week is a meaningful unit of change. A sprint review assumes the sprint is where the surprises are. A RAID log assumes risks arrive slowly enough to be logged by hand. None of those assumptions survives a team whose agents work nights.
What project managers actually do in this situation is what the one above does: reconcile. Read the merged work, guess which requirement each change was for, update the matrix, chase the two developers whose commits reference nothing. Wellingtone’s survey of PMO practice found 72% of practitioners spending half a day or more each month just collating reports55 — and that figure was measured before the volume doubled.
There is also a quieter failure. Forrester’s read of the agentic SDLC keeps the human accountable while AI does more of the execution.60 Accountability without a current picture of the work is not oversight. It is signing for a delivery you have not opened.
The obvious fix, and why it fails
The reflex is more reporting automation: a bot that summarizes the week’s commits, a model that drafts the status update. The summaries are fluent, and rooms stop trusting them fast, because a generated paragraph about velocity is exactly as checkable as the blended health score it replaced — which is to say, not at all. METR’s randomized trial is the cautionary number here: experienced developers using AI tools were 19% slower while believing they were 20% faster.61 Perception and record diverge precisely when the work speeds up. A summary is perception.
The alternative is to stop treating visibility as a document someone maintains and start treating it as a property of the record. A story that cannot exist without naming the requirement it serves. A requirement whose evidence is a file path pinned at a commit. A gate that opens only when a named person accepts. Then the Monday picture is not reconstructed from forty-one merges — it is already standing, because the work could not have been recorded any other way.
This is what we built BoardWalk around. The backlog is drafted from the requirement register in the register’s own words, so the trace exists from the first second rather than being drawn in retrospect by whoever owns the spreadsheet. A delivery card carries the requirement it serves on its face. When an executive asks how Broker Pathways is going, the answer is counts that keep their denominators — 119 requirements, 98 resolved, 0 verified by a named person — and every count opens to the rows behind it.
The project managers this serves are not being automated out. They are the ones the flattening promotes: when three engineers and a fleet of agents can build the thing, the scarce skill is knowing whether the right thing is being built, whether the evidence would survive an audit, and who has actually signed. That was always the job. The reporting was in the way.