The PR Description Your Coding Agent Generated That Humans Stopped Reading

A year ago your team adopted a PR description template. It had a ## Summary, a ## Changes, a ## Test plan, and a row of checkboxes. Reviewers loved it: every PR had context, every PR had a test plan, every PR had structure. Six months later the coding agent learned to fill it in. Now every PR has a ## Summary, a ## Changes, a ## Test plan, and a row of checkboxes — and reviewers no longer read past the title. The format that once focused attention now signals that there is nothing worth focusing on. The structure outlived the signal it carried.

This is not a code-quality problem. The code in those PRs is often fine. The problem is that the act of writing a description has been amputated from the act of thinking about the change, and the description is the artifact reviewers used to triage what to spend their finite attention on. When that artifact becomes uniformly formatted, plausibly worded, and indistinguishable from every other PR, the reviewer's attention triage breaks. The system that used to surface the unusual now flattens everything into the same shape.

The 2026 PR-description characteristics study found that Claude Code, GitHub Copilot, Cursor, and Devin produce descriptions with a recognizable shared shape: low use of varied headers, formulaic sectioning, and a verbosity that fills the template without prioritizing within it. The study tied that shape to lower readability scores. The lower-readability part is the one teams discuss; the more dangerous part is that the shape is also predictable, and predictable inputs are exactly what reviewer attention habituates to.

Habituation is doing the work, not laziness

Reviewers who skim agent PRs are not being negligent. They are doing what every nervous system on earth does when fed a repeating stimulus: they stop allocating attention to it. The clinical term is habituation. It is the same mechanism that makes the hum of a server room disappear from your awareness within minutes and the same mechanism that lets ICU nurses sleep through bedside alarms that fire dozens of times a shift without indicating anything. Habituation is not a character flaw. It is a feature of finite attention meeting infinite noise.

The pre-agent PR description used to be a low-frequency, high-variance signal. A given reviewer might see ten descriptions a week, each written by a different human, each shaped by the author's judgment about what mattered in this change. Some descriptions were two lines. Some were essays. Some called out a single risky line. The variance was the signal: the description's shape told you something about the change before you read the diff. A long description meant the author thought this one was tricky. A terse description meant the author thought the diff spoke for itself. The reviewer's eye learned to parse the meta-level shape.

Agent-generated descriptions strip that variance out. Every PR has the same four sections, the same paragraph rhythm, the same checkbox list, the same "Test plan" header followed by three plausible bullets. The shape no longer encodes the author's judgment because no author judgment shaped it. The reviewer's habituation kicks in faster than anyone planned: by month three of agent adoption, the description has the perceptual weight of the GitHub navbar, present but unread.

The halo effect is the second-order failure

Habituation drops the attention floor. The halo effect drops the suspicion floor. A January 2026 study found that AI-generated pull requests with nearly twice the code redundancy of human-written ones drew fewer negative reactions from reviewers — surface plausibility was reducing critical engagement on the code itself, not just on the description. The two effects compound. Reviewers stop reading the description because it looks the same as every other description. Then they reach the diff already primed to expect that the change has been thought about, because the description sounded thought-about. The well-formatted PR has had its quality-signaling channel hijacked: it now signals "skim and approve" not because it is bad, but because nothing in its shape distinguishes the changes that need scrutiny from the ones that do not.

A widely cited industry report this year put PR review time up 91% on teams with high AI adoption while merged PR volume rose 98%. Time and volume both went up, which means the per-PR review intensity dropped. Some of that is real efficiency. A meaningful slice of it is reviewers triaging by gut feel because the descriptions stopped helping them triage.

What a description was actually for

The literature on PR descriptions is consistent on what a good one does: it tells the reviewer where to look. It calls out the design decisions the author wasn't sure about. It flags the section that touches the gnarly part of the codebase. It marks the part of the diff that you would not understand from reading the code alone — the constraint that lives in someone's head, the customer incident from last month, the ambiguity in the spec the change resolves in a particular direction. The description is the author's gift of their context to the reviewer's attention budget.

An agent that has just produced a 600-line diff does not have most of that context. It can summarize what the diff does. It cannot reliably tell the reviewer what part the author was unsure about, because there was no author who held an opinion about uncertainty. It can list files changed, but listing files changed is not where the value lived. The agent-generated description fills the format and skips the function. The reviewer reading it cannot tell where to look, because the description was not written by anyone who knew where to look.

This is the deeper failure: the template was a contract between author and reviewer about how attention would be allocated. The contract had one side written by a system that does not have the information the contract was designed to convey.

The half-measures that do not fix it

Loading…