Each of the three defects named in the post turns on timing rather than syntax: a rate limiter whose check and consume were not atomic at the call site, a heartbeat that proved a worker was pinging rather than progressing, and a held database transaction that outlived the request that opened it [5]. An agent that writes the code and then writes the test for that code will pass all three [6], because the test inherits the same assumption about ordering that the code does. The observation worth keeping is not that agents write bad lines. It is that the repair does not travel: the patch lands in the file where someone noticed the symptom, and the class of defect has nowhere to live unless something outside the conversation records it as a standing rule [7].
That is why a stronger model is the wrong purchase for this particular problem. The author's own position is that a frontier model in 2026 writes fine syntax all day [13], and the bugs he documents were syntactically clean and green on their own tests [6]. Diff review, which is the thing most teams have actually instrumented, scores each turn in isolation and would have signed off on every one of them.
The remedy carries a price the post states in passing. LEO ships 41 codified laws and 22 specialist roles [1], which is 63 artifacts per repository that somebody has to write, keep current, and keep true [1]. The count is less interesting than the bookkeeping around it: the project's changelog logs every rule the system has ever added together with the reason it was added [4]. My read is that the reason field is the part that survives contact with a team, because a rule nobody can justify gets dropped the first time it is inconvenient, and a rule file that has quietly gone wrong instructs the agent with exactly the same authority as one that is still correct.
Be clear about the evidence. It is one practitioner's account: paid engagements on multi-tenant SaaS platforms, one of them with background AI pipelines [3], plus his own project's rule changelog [4]. No before-and-after defect counts appear, and no second team runs the rules against a codebase the author did not write [2]. The claim that the laws and roles stopped the recurrence is his [1], and it is the kind of claim only a comparison can carry.
The argument that generalises, whether or not LEO itself is any good, is about location. The author insists the protocols sit in files inside the repository the agent is working in, not in a paragraph of best practices somewhere in training data and not on a vendor's central server the team cannot see into [12]. Rules in the repo get reviewed, blamed and reverted on the same machinery as code. Memory held on someone else's server is a dependency you cannot diff.
