AINEWS
Your AI Agent Is Fast. Your Workflow Isn't.
Agents accelerate individual steps inside approval chains and handoffs that were never redesigned to move at agent speed, a pattern that echoes a century old factory mistake.
Agents accelerate individual steps inside approval chains and handoffs that were never redesigned to move at agent speed, a pattern that echoes a century old factory mistake.
A new report lands in a shared inbox at 9:14 a.m. An AI agent drafted the bulk of it in four seconds. The next action is human: a reviewer needs to read it, an approver needs to sign off, a calendar needs to clear. By Tuesday afternoon the report is four days old, and the total delivery time is roughly what it would have been if a person had written it over a long weekend.
The agent is genuinely good at its job. The system around it is unchanged.
That gap is the subject of a short essay by the operator and writer known as Sebas, titled "Electrifying the Cow Path". The argument is not that AI agents fail. The argument is that bolting them into existing workflows is a category error: the step gets faster, and the wall arrives sooner.
Sebas borrows the analogy from the late nineteenth century. Around 1900, factories began replacing central steam engines with small electric motors. A naive forecast predicted an immediate productivity boom. The boom did not arrive for decades, and historians of technology have spent a century explaining why. The early electric factories kept the layout that the steam engine had forced on them: a long central driveshaft running the length of the building, with belts reaching down to each machine. Wiring the building for electricity did not change the geometry. The motor moved to the machine. The machine did not move to the motor. Aggregate output was, for a long stretch, roughly the same as it had been under steam.\n Sebas attributes this historical pattern to economic historians Paul David and Robert Solow, whose research on the long delay between electrification and measured productivity gains provides the empirical backbone for the analogy.
The agents in today's offices face the same shape of constraint, Sebas writes. A report can be drafted in seconds, but the queue it lands in was sized for reports that took days. Approvals, handoffs, human-in-the-loop checkpoints, and the calendar slots that hold them are still tuned to the old latency. A faster first step does not shrink the wait. It only changes what people are waiting for. In queueing terms, the arrival time shortens while the service time does not, so the system-level wall moves closer to the start of the chain and gets there sooner.
The piece is an opinionated analysis rather than a measured benchmark, and the historical productivity pattern it draws on traces to economic historians Paul David and Robert Solow, who first documented the long lag between electrification and any pickup in measured output. Sebas does not produce a new number. He uses the historical record as a diagnostic for what is happening now. The cow path in the title is the workflow as it is, the path worn into the ground by years of human pacing. Electric motors on a cow path still produce a cow path.
The constructive move, and Sebas is explicit about this, is to point at the actual unlock. The agent stays. The chain gets redesigned: reviewers read exceptions instead of full drafts, approvers set the rules an agent is allowed to operate under, and calendars stop being sized to the old latency. The win comes from changing the geometry of the work, not from installing a faster motor on the same shaft.
The test for any deployment is small and uncomfortable. If a faster step inside the chain leaves the surrounding handoffs, approvals, and human-decision moments exactly as they were, the savings will land on the floor of the queue and stay there. The work of agent rollouts is not the agent. The work is the path.