Coverage at 80% was a price on human attention, and CI is the wrong place to charge it

The direction the feedback travels is the load-bearing part. A linter that fails in CI thirty minutes after the agent has stopped is addressed to a person, not to the thing that wrote the code ^[5]. The post's test for admission to the harness is one command the agent can run itself, failing loudly enough that it cannot be ignored ^[6]. That demotes a familiar piece of infrastructure. Most teams call CI the enforcement layer; on this reading it is the reporting layer, and enforcement has moved to whatever the agent can invoke before the branch exists ^[5].

Then the arithmetic. Moving from the industry's 80% to a 100% floor shifts one branch in five out of optional and into mandatory ^[1]. That fifth is not a random sample. The author's point is that an agent skips the awkward branch precisely because it is the annoying one to set up ^[7], so the portion everybody wrote off is where the fixtures, the doubles and the error paths live. It was never the cheap 20% left over. It was the expensive 20% that human review time could not afford.

On an existing codebase the 100% minimum applies only to files a pull request touches, so coverage ratchets upward as the agent works through the code ^[8].

Coverage remains the weaker of the two checks described here. Left alone, an agent adds defensive code that does nothing, the classic being an emptiness check before looping over an array, when looping over an empty array already does nothing at the same speed ^[9]. Tests execute that line, so coverage marks it covered. Mutation testing is what catches it: change the line, and no test cares ^[10]. A suite sitting at 100% can still be certifying noise, which is worth knowing before anyone treats the number as an outcome.

Two of the failure modes are specific to generated code. A model working from a half-remembered API emits a method that looks entirely plausible and does not exist, and a type checker rejects it statically with a precise error rather than a stack trace the agent has to interpret ^[11]. And a model's instincts come from its training data, so it reaches for the idiom that was correct in 2019 with complete confidence; the example given is Time.now in a Rails codebase where Time.current is the timezone safe call, a bug rather than a style preference, and one an agent will reintroduce every few weeks unless something stops it ^[12]. Insecure output follows the same pattern: the mistakes are learned from existing codebases, which means detectors for them already exist ^[13].

This is one practitioner's account, and the toolkit is Ruby-shaped, with rspec, simplecov, mutant, sorbet and rubocop named rather than any before-and-after measurement ^[14]. The mechanism is the portable part. Any threshold your team justified by the cost of looking at things is now a setting, and someone should reopen it ^[4].