Notes on systems & accountability
AI Efficiency and Merge Gate Design
Why agentic merge gates are efficient during a build-out, and why the bottleneck should sit at ready-column intake rather than on every pull request.
A software build-out is a period of high change volume. The useful property of an agentic merge path in that phase is not that review has been solved. It is that completed work can reach the main branch quickly enough for the next change to land on a known state.
That speed has a real design purpose. Small, frequent merges reduce the cost of integration. They keep conflicts short, keep tests running against current code, and let a later change begin from what actually shipped rather than from a stack of unmerged branches. A merge gate that waits for a person to read every diff will not merely slow publication. It will turn the merge queue into the planning queue, and the build will start organizing itself around what can get through a scarce reviewer instead of around what should be built next.
The structural imbalance is straightforward. Agents can open pull requests faster than a person can review them with care. If the official gate is a human reading each change, two failures follow. Either the queue grows, or review becomes ceremonial: a glance, a green check, and a merge that no longer represents skepticism. Neither outcome is a review process. One is a backlog. The other is a rubber stamp with a human name on it.
There is a tempting workaround: let another agent review the first agent’s work. Automated checks belong in that loop. Tests, linters, acceptance contracts, and a second pass over the diff can catch a large class of errors at machine speed, and they should. The risk is treating a second agent as a substitute for human skepticism. Independence is not created by a new session, a new model, or a new summary. Two agents can share the same narrow test, the same assumption about what “done” means, and the same willingness to accept a plausible report. A closed loop of confident reviews can pass work that a skeptical reader would still question. Agent-on-agent review is a useful filter. It is not, by itself, a gate.
The gate that should stay tight is earlier. The bottleneck belongs at ready-column intake, not at human review of every pull request. Ready should mean that a person, or a clearly owned process, has decided the work is specified, unblocked, and worth doing. After that decision, agents can implement and merge under automated checks. Human attention then goes where it changes the work: choosing what enters ready, verifying outcomes that matter, and handling exceptions. It should not be spent as a per-request merge tax.
A companion note, Merge Velocity: Where Work Stalls, reports one week’s figures. In the window from November 7, 2026 through 9:30 AM ET on November 14, 2026, 228 pull requests merged, about 29 a day, with a median open-to-merge time of about 3 hours 20 minutes and with 90 percent merging within 15.3 hours. The review queue was not growing: 8 open pull requests at the end of the window, after swinging between 4 and 14 since November 11, and issues in review down from about 47 on November 9 to 34, most of them already merged and waiting on an owner verification line. The ready column held 1 issue, against 22 blocked and 31 parked. Those numbers do not prove that intake was deliberate. They do show that merge speed and review depth were not the constraint, and that the shortage was already in the supply of ready work.
That is the shape a build-out gate should have on purpose. Be strict about what becomes ready. Be fast, and mechanically checked, about what then merges. Keep a human verification step for outcomes that the merge itself cannot prove. And do not confuse a short open-to-merge time with a measured time-in-column: the earlier note states plainly that column dwell time is not yet being logged.
Efficiency in this phase is not unbounded throughput. It is the decision to place scarce human judgment where it selects work, and to refuse to use that judgment as the thing that meters every merge.