The output can scale before the finish line does.
Anthropic's study of its own workplace reports that employee-directed agent runtime reached six times estimated human work hours by August 2026. The report also estimates that reading interactive agent output would take more than half a work week for non-technical groups. Its authors describe human attention as a bottleneck and report that employees spent more sessions building instructions, monitors, and agent-management systems.[1]
This is one AI company studying itself. Its runtime estimate excludes waits longer than 30 minutes, while human time assumes a 40-hour week. Its reading estimate converts tokens to words and assumes 250 words per minute. Use the report as a detailed case, not a universal staffing ratio.
Do not add an agent until you know what leaves the review queue.
More commits do not become releases at the same rate.
A 2026 NBER working paper studies more than 500,000 GitHub developers with matched AI usage data. The authors estimate that autonomous coding agents increased commits by 240 percent, projects by 80 percent, and releases by 30 percent. They interpret the drop across those measures as evidence of human bottlenecks in production.[2]
The paper measures broad populations and reports associations from its study design. It does not tell one team how long a review should take. Its useful warning is narrower. Count the output at each production boundary. A commit, project, release, and used product are different units.
Comprehension checks can protect one narrow lane.
Sphinx is a new Git hook that pauses configured pushes and asks the operator questions about the committed change. Its documentation says the reviewer reads through read-only Git tools, does not see the conversation that produced the code, and can be bypassed by anyone with local access. The project calls the check practice rather than a security gate.[3]
The published sphinx-understand package existed at version 0.1.4 during this pass. Its isolated help command exposed install, init, status, check, and doctor commands. We did not install hooks, connect a model provider, or run a live comprehension check. A quiz can reveal missing understanding. It cannot prove runtime behavior or approve a release.
Give review a work-in-progress limit.
- Estimate review minutes per change from recent work, not the fastest clean patch.
- Reserve review hours before choosing how many agent tasks may run.
- Cap the number of changes waiting for review.
- Stop launching work when the queue reaches that cap.
- Remove stale drafts instead of letting them compete with current work.
- Track accepted changes, reversions, and releases beside draft volume.
The load bank below turns those inputs into a teaching estimate. It does not read your repositories, measure change complexity, or forecast defects.