--- title: 'In the News: October 28, 2026 (Evening)' description: "GitHub says over 40% of pull requests now start with an agent's draft; Snowflake's monorepo benchmark shows task wording moved one pass rate by 26 points." canonical_url: 'https://darkfactory.dev/news/2026-10-28-evening' markdown_url: 'https://darkfactory.dev/news/2026-10-28-evening.md' collection: news date_published: '2026-10-28T19:05:00-04:00' date_modified: '2026-10-28T19:05:00-04:00' --- # In the News: October 28, 2026 (Evening) GitHub's annual report puts agent-started pull requests above 40% of the platform's total, and its merge-rate data says reviewers are accepting them at least as often as other pull requests. Snowflake's engineering blog adds a measurement from the other side: on one task, the wording of the description moved agent pass rates more than any difference between harnesses. ## 1. GitHub: more than 40% of pull requests now start with an agent's first draft **[Octoverse: Commits, pushes, reach new highs in 2026 as developers build with agents](https://github.blog/news-insights/octoverse/octoverse-commits-pushes-reach-new-highs-in-2026-as-developers-build-with-agents/)** · GitHub Staff · The GitHub Blog, October 28, 2026 GitHub reports that by September 2026 more than 40% of pull requests started with an agent's first draft, up from 17% in March. That is 110 million in September alone, about six times March's volume. Agent-started pull requests changed multiple files more often (about 87%, against 79% for pull requests without detected agent involvement). In September 90% of agent pull requests merged within two weeks, against about 86% for the comparison group; in February the agent figure was 83%. GitHub says its agent detection is "a floor", so the true share is likely higher. It also reports that repositories created this year with an agent configuration file logged eight times as many commits as new repositories without one, and that Copilot cloud agent, where a developer assigns a task and the agent works independently, grew 230% year over year against 85% for agent chat. Review tooling grew alongside. AI-assisted code review usage grew more than 90% year over year, and more than 264,000 developers had used stacked pull requests, which launched as a preview on July 30. GitHub sells Copilot, and all figures are its own telemetry; the post does not report defect, revert or incident rates. **Why it matters:** Merge rate cannot separate good agent output from lightly reviewed agent output, and this report measures only the first. If a team's review capacity is flat while agent-started pull requests triple, the merge numbers will look healthy either way, so a team should track what happens after merge, not just whether review approved it. ## 2. Snowflake: the task description moved pass rates more than the harness did **[SnowPR: Benchmarking Coding Agents on Snowflake's 43-Million-Line Monorepo](https://www.snowflake.com/en/blog/engineering/Benchmarking-coding-agents/)** · Soyoung Yoon and five co-authors, Snowflake · Snowflake Engineering Blog, October 27, 2026 Snowflake built a 65-task benchmark from its own merged pull requests, graded by the tests that shipped with each change and run on its own Bazel build farm. With opus-5, four harnesses passed 84.1% to 85.1% of trials, a spread the authors say 65 tasks cannot detect. Cost per trial differed: $6.34 for Cortex Code in its REM mode, $6.51 for Claude Code, $6.93 for Cursor and $8.19 for Cortex Code in code mode. Changes that cross the Java and C++ boundary passed 55% of trials, against 80% for Java alone. For one pull request, the same hidden tests and reference fix produced an 83% pass rate on one engineer's write-up of the task and 57% on another's, across 10 agents. The authors also report that an LLM judge's predicted pass rates "barely tracked" measured ones (Spearman +0.18 over 51 tasks), so they use the judge only to filter candidate tasks, not to rank results. Snowflake builds Cortex Code, one of the harnesses compared, and REM mode is internal only for now. We found no link to the task set in the post. **Why it matters:** The post measures one pull request with two descriptions, so the 26-point gap is a single data point. Even so, it is larger than any harness gap the authors could detect, which suggests a team comparing harnesses should fix the spec wording first, then compare cost per trial. --- ## Also this cycle - **[github/lever](https://github.com/github/lever)** · GitHub · A Copilot plugin of skills for setting up feedback loops. `/lever loop` returns a "Loop Card" listing the metric, how to measure it, how noisy it is and what counts as a win, and an autonomy setting runs from L0, where the agent only proposes, to L5, where it lands, releases and watches. The README calls the project early and reports no measurements; the repository had 52 stars when read. - **[Software Factories for Real Engineers](https://www.aihero.dev/cohorts/software-factories-for-real-engineers-xdy1m)** · Matt Pocock, AI Hero · A two-week paid cohort starting November 9, listed at $1,295. The sales page argues teams should build brakes (automated checks, automated and human review) before accelerators (schedules and triggers that start agents unattended), saying "starting with accelerators is backwards." The lessons are unreleased, so this is the seller's description of material we have not seen.