Harness moves the code review gate off the reviewer's calendar

A pull request sits for two days because the only reviewer who knows that service is in planning sessions all week. That gap is what Harness is selling against, and it pays to name it precisely: queue time, not thinking time. Chief Executive Jyoti Bansal's description of the status quo is a queue by design, with a human opening a pull request and colleagues tuning and approving over hours or days ^[3]. It worked while the inflow was also measured in days ^[14].

The one number on offer is 10,000 hours saved in a month, self-reported by Harness engineers using Harness tooling ^[10]. At a 160-hour working month, that is about 62 people's worth of time ^[13]. It is also a figure without a denominator: no team size, no baseline review latency, no defect or revert rate to set the hours against ^[16]. Hours saved has no failure mode, which is part of why it is the number that gets published.

Bansal says the entire software delivery lifecycle has to become autonomous, with repository, review, pipeline and governance running as one system ^[11]. The shipped thing is narrower and more useful than that sentence: a queue that does not block when thousands of pull requests land at once ^[2], and a check that runs at merge and pushes failures back ^[7].

The permission model is the design choice worth reading twice. Inheritance is cheap to build and honest about accountability ^[5], and here is what it means on a Friday: if one engineer's credentials launched forty agents overnight, that engineer's name sits on every merge those agents made. The audit does not get an agent to blame.

For deciding which checks to make mandatory, two axes do most of the work. First, whether a rule can settle it, such as tests passing or a dependency policy, or whether it needs judgment about whether this is the right abstraction. Second, what a miss costs, a revert or a customer. Rule-settled and expensive is where mandatory automated checks belong. Judgment and expensive stays with a named human and gets scheduled rather than queued, because queueing it is how you got here. Rule-settled and cheap can be advisory. Judgment and cheap is a comment somebody should stop writing.

Harness is explicit that a human still decides what ships, and puts AI Code Review at the gate to make that call faster rather than to remove it ^[9]. That is the honest version of the pitch, and it sets the metric too. Over the first sixty days, the useful number is not hours reclaimed but the revert rate on agent-authored merges, because that is what tells you whether the checks you made mandatory were the ones that mattered. What teams tell themselves is that review is where design judgment happens. What the timestamps in most repositories show is elapsed time spent waiting for attention, and any tool that only speeds up the reading will leave that untouched.