The strictness in that merge queue comes from what the reviewing agent lacks, not from any upgrade. An agent that helped build a branch carries the reasons the branch is fine. An agent that meets the same branch at a gate has the repository and the rules and nothing else, so the rules are the only thing it can check against, which is what the cate.blog post credits for the stricter review [8]. Its author says she has given up on metaphors and treats that contextlessness as a neutral property rather than a strength or a weakness [16]. Operationally it behaves like a dial you get to set. Inside the work, an agent moves fast and forgives its own decisions. Outside it, the agent enforces what you wrote down and nothing you did not.
That is where the documentation bill lands. Scaling has always meant converting tribal knowledge into documentation, automation and process; only automation used to be written for a machine [1], and now all three are [1]. The doc that was a courtesy to a teammate is an input to a work product, and the process that lived in someone's head has to be parseable [1]. With people, knowledge sits in the people and in the systems; with agents, only the systems half survives the session boundary [2].
The test I would put on a backlog item tomorrow has two axes. Axis one is whether the knowledge is written somewhere an agent actually reads. Axis two is whether it is checked at a gate the work cannot route around.
- Written and gated: the merge queue, which serialises merges, files follow-ups and returns PRs that are not ready [6]. - Written, not gated: the encoded instruction to dispatch a review agent, which helped for months and still leaked the odd issue and untracked follow-up [5]. - Gated, not written: a rule that blocks work for reasons nobody on the team can reconstruct. Cheap to add, expensive the first time it is wrong. - Neither: the tribal knowledge, re-derived from scratch each session, in ways you find out about on Friday.
The human pacing rule was one careful change a month, on the reasoning that no effective person wants a job that feels more like process than impact [12]. That constraint was about people's tolerance, and it does not govern an artifact a machine reads. The post's own example is CLAUDE.md as a living team document whose churn is treated as a good sign [10]. On an earlier platform, the team moved off the affectionate "intern Claude" framing, put guardrails in, and started treating the thing as a system to be designed rather than a novelty to be supervised [9].
The uncomfortable part is the one the author names herself: people will do for a machine what they were never willing to do for their team, and she declines to claim moral high ground over anyone for it [11]. Automation used to arrive after an incident, because the incident was what made the cost of not doing it legible to everyone who had to approve it [13]. Now a contextless agent supplies that same argument every morning, for the price of one wasted session.
