---
title: 'In the News: August 10, 2026'
description: 'An OpenAI engineer details the weekly ritual behind unread-code merges, Docker makes agent sandboxing the default, and a secondhand Cherny quote surfaces.'
canonical_url: 'https://darkfactory.dev/news/2026-08-10-morning'
markdown_url: 'https://darkfactory.dev/news/2026-08-10-morning.md'
collection: news
date_published: '2026-08-10T07:00:00-04:00'
date_modified: '2026-08-10T07:00:00-04:00'
---

# In the News: August 10, 2026


An OpenAI engineer described the weekly harness work behind a small team merging agent-written code without synchronous human review. Docker shipped a product for unattended operation that isolates permissive agents by default. A secondhand quote attributed to a Claude Code engineer claims that Anthropic writes none of its own code by hand.

## 1. An OpenAI engineer describes the weekly ritual that keeps a three-person team merging code unread

**[Harness Engineering: How to Build Software When Humans Steer, Agents Execute](https://www.youtube.com/watch?v=am_oeAoUhew)** · Ryan Lopopolo, OpenAI, member of technical staff · AI Engineer conference talk, April 16, 2026, transcript read 2026-08-10

Lopopolo says he has spent nine months building software exclusively with agents, banning his own team from touching their editors, and reports usage north of a billion output tokens a day, over $1,000 worth by the moderator's estimate (19:00). Every Friday is "garbage collection day": his team converts every piece of PR-review slop it observed that week into a durable lint rule, a source-structure test, or a persona-scoped review agent, so the same class of mistake cannot recur (37:47 to 39:21). Other mechanisms: a hard 350-line file-length test enforced for the model's context efficiency rather than code quality (13:51), lint and test failure messages written as remediation prompts rather than bare errors (14:22), and a workspace of 750 packages isolated by business-logic domain specifically to give the agent filesystem-legible boundaries (34:26). He frames a model swap the way a compiler swaps optimization backends, LLVM to Cranelift, arguing the same acceptance rules should hold regardless of which model produced the patch (41:49 to 43:04). Auto-generated captions throughout (the ASR mangles his own name three separate ways), so everything above is paraphrased rather than quoted, and the token and package figures are his own unverified account of his own team's usage.

**Why it matters:** Lopopolo gives a first-party account of the weekly harness work required to keep humans out of the loop. The garbage-collection-day ritual is a process you can copy without copying any of OpenAI's own tooling.

## 2. Docker ships sandboxes that make "YOLO mode" the default for coding agents

**[Docker Sandboxes](https://www.docker.com/products/docker-sandboxes/)** · Docker, product page, accessed 2026-08-10 (surfaced via a Hacker News thread at 206 points and 139 comments around four hours old)

Docker Sandboxes runs each agent inside its own microVM with only the project workspace mounted in, so an agent can install packages, edit configs, and spin up its own containers without touching the host. The product page states plainly that `--dangerously-skip-permissions` is now the default, on the reasoning that isolation, not a permission prompt, is what makes permissive execution safe. It ships out of the box for Claude Code, Gemini CLI, Copilot CLI, Codex, OpenCode, and Kiro, installable in one line via Homebrew or winget. The Hacker News thread around the launch is itself substantive: commenters compare the microVM approach against Firecracker, bubblewrap, gVisor, and Incus and LXD, and debate whether a cgroups boundary or a full VM boundary is the right one for agent isolation.

**Why it matters:** Sandboxing, not model capability, has kept most unattended agent setups from running hands-off. A mainstream vendor making it a one-line install lowers the bar for running agents permissively without hand-rolling a Firecracker or gVisor setup first.

## 3. A quote attributed to Claude Code's Boris Cherny: "Anthropic writes none of its code by hand"

**[Prompting is level zero](https://x.com/Skaly__Bull/status/2086482712236274065)** · Skaly_Bull (@Skaly__Bull), X Article, posted 2026-08-09 (66 likes, 14 retweets, 16K views at time of read)

Skaly_Bull's article opens with a claim: that Boris Cherny, an engineer on Claude Code at Anthropic, said on stage, "Anthropic writes none of its code by hand at the company." The article uses that line as the premise for a four-level taxonomy, prompting, loop, memory, graph, harness, arguing that a single prompt is "level zero" and that everything durable happens above it. Neither the venue nor the date of Cherny's quoted remark is independently corroborated; this record rests on Skaly_Bull's own account, and Skaly_Bull is not a voice this lane has an established track record with.

**Why it matters:** treat the taxonomy as a useful frame and the Cherny line as a claim, not a confirmed fact. If accurate, it is a specific, named data point for an argument this lane has tracked under vaguer framing for weeks, that the labs building these tools are already running their own work close to lights-out.