{"slug": "the-cold-read", "title": "The Cold Read", "summary": "A practitioner running roughly 200 pull requests a day has added a \"cold\" review agent that receives only the diff, with no PR description or prompt, and describes the change; a second judge agent then grades how well that description agrees with the original on an A-through-F scale. The cold read feeds the assessment deciding whether a PR can merge itself, so a description the diff does not back up counts against automerge; in one case it flagged a PR described as a performance enhancement whose memory cost had not been assessed. The author argues the approach works because the cold agent shares only the diff with the agent that wrote the code, making it an independent second view of the change.", "body_md": "I give every PR to a separate agent \"cold\". It gets the diff alone, with no description and no prompt, and I ask it to describe the change. Then a second agent judges how well that description agrees with the original (PR description, prompt, etc).\n\nA PR description is a *claim*<sup>[\\[1\\]](#fn1)</sup>. On our PRs the agent that wrote the code also writes the description, and it's the first thing a reviewer reads, human or model. So every review starts out already told what the change is<sup>[\\[2\\]](#fn2)</sup>.\n\nHere's one it caught. The description on a PR said it was a performance enhancement, and in one way it was. But there was a cost (memory) and that had not been assessed. The usual reviews had read \"this is a perf enhancement\" and through that lens it looked true. The cold agent hadn't read it, and its account of the diff didn't match.\n\nIt sits alongside the [normal review](https://jonathannen.com/my-current-ai-workflow-revisited/) and runs on every PR, which is around 200 a day. It's the same idea as the separate runner in [Tunnel Vision](https://jonathannen.com/tunnel-vision/), a fresh context that wasn't there for the original task.\n\nThe judge grades how well the two agree, A through F. It tells me how much I can rely on the description.\n\nIt feeds the [assessment that decides whether a PR can merge itself](https://jonathannen.com/start-with-one/), so a description the diff doesn't back up<sup>[\\[3\\]](#fn3)</sup> counts against automerge. I eyeball it too. When the two accounts agree I can merge by hand on the strength of the description, and when they don't I read the diff.\n\nIt has made automerge more reliable, and I trust it more. I argued in [Views of the Same Thing](https://jonathannen.com/views-of-the-same-thing/) that a second view of a change only pays its way when it comes from an independent reading. The cold read is one, since all it shares with the agent that wrote the code is the diff.", "url": "https://wpnews.pro/news/the-cold-read", "canonical_source": "https://jonathannen.com/cold-read/", "published_at": "2026-11-13 00:00:00+00:00", "updated_at": "2026-11-18 21:41:37.650920+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "developer-tools", "ai-safety"], "entities": [], "also_reported_by": [], "takeaway": "Run a separate agent on the raw diff alone, with no PR description, and have a judge grade its account against the description A through F; a mismatch should block automerge, as it did for a PR billed as a performance enhancement whose memory cost went unassessed.", "recommendation": {"advice": "Give a fresh agent only the diff, no description or prompt, and grade its account against the PR description; block automerge when they disagree, since the author's cold read caught an unassessed memory cost at 200 PRs a day.", "playbook": "coding-agents", "playbook_url": "https://wpnews.pro/playbooks/coding-agents"}, "alternates": {"html": "https://wpnews.pro/news/the-cold-read", "markdown": "https://wpnews.pro/news/the-cold-read.md", "text": "https://wpnews.pro/news/the-cold-read.txt", "jsonld": "https://wpnews.pro/news/the-cold-read.jsonld"}}