The authors built RepoPremium, a file-ranking benchmark whose core has 1,265 queries over ten SWE-bench Verified repositories. It pairs the cleaned-up issue statements SWE-bench uses with raw pull request text from 2022, 2023 and 2026,…
In the News: November 21, 2026 (Extra)
An Ant Group code-search study finds much of the embedding lead over keyword search comes from how benchmarks are set up.
Extra edition
Machine-readable
Download Markdown
Story
Much of the embedding lead in code search comes from how benchmarks are set up
Read story →
An expedition for expeditious PR reviews
LaunchDarkly's Agent Platform team submitted nearly twice as many PR reviews from late September to late October as from late May to mid-June, which Brackbill links to coding agents writing more code. Auditing two weeks of PRs in the company's main repository, he counted an average of 13 notifications per approved PR, mostly CODEOWNERS pings for cross-cutting changes. The team now has agents stage review comments in a pending GitHub review, asks them to keep comments within 300 characters, and uses a round-robin for CODEOWNER duty so only one or two people get each ping.