Wes McKinney, the programmer who created the pandas data library, says his three-person startup merges hundreds of pull requests a week while keeping bug counts low across a codebase spanning millions of lines. He credits none of it to letting coding agents run unsupervised. The mechanism he describes is a review pipeline built to catch mistakes before they land, not a faster way to generate code.
McKinney runs Kenn Software, a startup building developer tools that has not yet launched its product. In a post on his personal blog, he pushed back on the current wave of “loop engineering” and “graph engineering” pitches circulating on social media, arguing that fully autonomous pipelines without a human in the loop produce unreliable output. His team’s throughput, he writes, comes from the opposite instinct: more checkpoints, not fewer.
The workflow starts with humans, not agents. Every design decision runs through a person before it hardens, and unclear calls get a second opinion from an agent session running a different model family than the one doing the implementation. Only after a design is settled does an agent turn it into a written specification, which a separate agent then reviews adversarially, with the cycle repeating until the review stops surfacing problems.
Implementation happens in small increments, checked by roborev, a continuous local code verification tool McKinney’s team built and now also sells. Every roborev finding on a branch gets closed out before a plan counts as done, and McKinney says the newest frontier models still produce work that is “extremely sloppy” and unfit for production without that hardening pass, sometimes costing hundreds of dollars in tokens per changeset to fix. A human always writes the pull request description and owns the merge decision; the team’s internal rules explicitly ban agents from merging without authorization.
To support that pace, Kenn built its own tooling rather than adapting GitHub’s stock workflow: Forge for reviewing and landing changes, Ghosthub as a terminal built for remote and multiplexed agent sessions, Kata as an agent-native issue tracker, and AgentsView for tracking token spend and agent behavior. McKinney cites his own usage as an example of the volume involved: at API list pricing, his agent consumption over one recent 30-day span would have cost $56,836, an amount subscription plans currently absorb.
None of this is independently verified. The PR count and the bug rate are both self-reported, and a PR count in particular is a metric a team can inflate simply by splitting changes into smaller commits. McKinney does not publish a bug-rate methodology, an incident count, or an outside audit, so what he has published is a practitioner’s account of his own team’s process, not a study with a control group.
That distinction is still useful. AI Insiders is covering a separate argument this week that verification, not raw model capability, is now the binding constraint on agent adoption. McKinney’s account is a real-world data point for that thesis: his team’s output scaled because the review discipline around the agents tightened, not because the underlying models got meaningfully better at writing correct code unsupervised.
Engineering leads evaluating agentic coding tools should treat Kenn’s numbers as a floor for what review infrastructure a serious deployment requires, not as a benchmark to match. The transferable part is not the PR count; it is the adversarial spec review, the branch-level verification pass, and the rule that a human signs every merge.
Wes McKinney, the creator of the pandas data library, described the workflow in a post on his personal blog, wesmckinney.com, published in August 2026.
