AI review as first-class CLI objects, not browser tabs
Harness CLI 3.0 (Sep 9) puts risk scores, review groups, and check→pipeline log streaming in the terminal so agents can triage and close the review loop without the 12-tab dance.
Source: AI Code Review Loop in the Terminal: Introducing Harness CLI for Harness Code
Mohit Suman’s Harness CLI write-up is the complementary surface to yesterday’s PR-flood pitch: same platform, but the load-bearing claim is that review insights are objects you can get, not comments dumped into a browser. harness get pr:insight and pr:review_group return risk tiers and logical change buckets before you (or an agent) reads forty files alphabetically. Checks carry pipeline and stage IDs, so get pr_check:log streams the failing stage instead of opening a second authenticated UI. Cross-repo inboxes (pr:review_pending, pr:mine) collapse the “which repos are waiting on me?” search into one command.
The agentic part is deliberate. The CLI is self-describing (list noun --matrix, get noun pr --format json), so an agent discovers verbs at runtime instead of hallucinating an SDK. The copy-paste autonomous loop is the same shape every software factory eventually needs: triage by AI risk group, stream failing CI, fix, thread a reply, approve/merge — all with JSON output and scoped tokens. Humans still get --ui hotkeys; agents get deterministic grammar and structured review objects.
Why it matters for code review
Agent PRs do not just raise volume; they punish context-switching. If review risk, ownership hints, and CI failure evidence live only in web UIs, the harness cannot close the loop. Making those signals CLI-native is how you spend human attention on high-risk buckets and let agents burn tokens on the rest.