---
title: 'In the News: November 9, 2026 (Extra 3)'
description: "A build tool claims 12 agent worktrees compile in 9.3 seconds against Cargo's 3 min 42 s. A pattern language says gates, not reviewers, should judge AI code."
canonical_url: 'https://darkfactory.dev/news/2026-11-09-extra-3'
markdown_url: 'https://darkfactory.dev/news/2026-11-09-extra-3.md'
collection: news
date_published: '2026-11-09T23:04:14-05:00'
date_modified: '2026-11-09T23:04:14-05:00'
---

# In the News: November 9, 2026 (Extra 3)


Giving every agent its own worktree means every agent rebuilds the same dependencies, and one engineer has measured that bill and built a cache to avoid it. Separately, a new pattern language argues that AI-written code should be judged by checks the author cannot influence, with people no longer reading diffs.

## 1. A Rust build cache for agent worktrees reports 9.3 seconds against Cargo's 3 min 42 s

**[Cargo Takes Minutes. Mitts Takes Seconds: 12 rustls Worktrees](https://prod.codes/blog/agents-rebuilding-the-same-crates/)** · Alexander Panasenko · prod.codes, November 9, 2026

Panasenko, who built the Mitts build tool, reports that a parallel build of 12 rustls worktrees on Linux x86_64 took 3 min 42 s with plain Cargo and 9.3 s with Mitts. Cargo with a warm sccache took 3 min 59 s. The shared caches were prewarmed, the local target directories were empty, and the warm-up build is excluded. On a Mac, he reports that twelve actix-web worktrees took 5 min 22 s with Cargo, 5 min 23 s with sccache and 56 s with Mitts, with 3,408 compiler invocations for Cargo against 181 for Mitts.

He also tested the obvious shortcut. A shared target directory across concurrent builds at different commits failed his check: an injected test failure in one worktree was missed or leaked into a clean one, so the result was excluded from the scores. Mitts instead shares immutable, hashed build results and keeps each worktree's output separate.

The post is open about where Cargo wins. Plain Cargo finished a single cold build first on 71 of 100 Mac projects and 60 of 90 Linux x86_64 projects, and after a one-file edit Cargo still won 22 of 71 Mac projects. Each reading is a median of three repeats, which he says is not a confidence interval. Mitts is pre-release, supports Linux and macOS only, and has no remote or team cache.

**Why it matters:** If your agents each work in a worktree, repeated compilation is a cost you can measure on your own project before you pay for more machines. The numbers come from the tool's author on a corpus of 100 open-source Rust projects, and we have not rerun them.

## 2. ZTHR proposes a pattern language where gates replace human diff review

**[ZTHR: Zero Trust, High Reliability](https://zthr.ai/)** · Russ Little · zthr.ai, v1.0, CC BY 4.0 (DOI 10.5281/zenodo.23826734)

Little's executive summary argues that AI has made writing code cheap but has not made deciding whether it is right cheaper, and that human review under real volume becomes either a bottleneck or a rubber stamp. His answer is a factory of checks: "Let AI write the code, and build a factory that cannot be fooled about whether the code works." No author, reviewer, model or report gains authority by asserting success, and a test that has never been shown to fail is not trusted to pass.

The summary says every change is reviewed by at least three models from different families before commit: one given the context, one given only the diff, and one checking the code independently. A fourth reviews it in place before full promotion. When a defect slips through, a new check is added so that kind of defect cannot pass again. People define intent and govern the checks; destructive or multi-step changes, any change to the AI itself and any agent request for wider permissions still need a person's sign-off.

The author says ZTHR is in production use for greenfield systems in a regulated industry, where it "has proven highly effective." The summary gives no figures for that claim, and we read only the executive summary, not the five-page brief or the full paper with its thirty-three patterns.

**Why it matters:** The design question is concrete: what evidence does a change need that its author cannot supply? The multi-model review rule and the "new defect, new check" loop are specific enough to test in your own pipeline. The effectiveness claim is the author's own and unmeasured in what we read.