---
title: 'In the News: November 19, 2026 (Extra 2)'
description: 'A software factory that has merged 1,940 PRs since August now lets its agents ask the human about six kinds of decision. Nothing else.'
canonical_url: 'https://darkfactory.dev/news/2026-11-19-extra-2'
markdown_url: 'https://darkfactory.dev/news/2026-11-19-extra-2.md'
collection: news
date_published: '2026-11-19T16:05:00-05:00'
date_modified: '2026-11-19T16:05:00-05:00'
---

# In the News: November 19, 2026 (Extra 2)


Calvin Grunewald has published the build log of the software factory behind his company's monorepo, including the six kinds of decision its agents may still bring to him. Separately, Bespoke Labs measured whether teams of eight agents gain anything from messaging each other. On average, they did not.

## 1. A working factory limits its questions to the human to six kinds

**[A Look Inside the Factory](https://thebuildingmachine.com/p/a-look-inside-the-factory)** · Calvin Grunewald, ArchAstro · The Building Machine, November 19, 2026

Grunewald built a harness, archdev, in June and added a factory mode in August. An overseer agent plans Tasks from a goal he approves and hands them to workers, each coding in its own worktree. Workers push to a private bare repository on his machine. A daemon turns each push into a durable SQLite job and runs lint, tests, policy checks and multi-agent review against the exact commit before anything reaches GitHub, then fixes review comments, CI failures and merge conflicts on the open PR. He reports that "Since August 21, the factory has merged 1,940 PRs into our monorepo."

Once the factory ran around the clock, the overseer kept blocking on him. By September 30 it had paused Tasks 67 times without telling him, and two days earlier workers had sent it 151 decision requests in about 21 hours, many of which they could have settled themselves. A question is now a paused Task with a kind, a reason and a recommended answer, and it must name one of six categories or the overseer rejects it: "The six are credentials, spending, live infrastructure (merging counts), irreversible actions, product decisions and anything the Task says needs my sign-off." For anything else the overseer decides and records why on the Task.

Review policy sits in a per-repo `archdev.json` rather than in the daemon, so it can be replayed against old PRs. On October 20 he ran six variants over 19 graded cases. None fixed more findings than the current config. A cheaper model cost about 3.5 times less per fixed finding, but on one case it deleted the public wrapper a reviewer had complained about. All figures are his own, from one team's monorepo, and archdev is his company's own tool.

**Why it matters:** The six-category list is a policy you can adopt without his tooling: make an agent name which human-owned decision it is blocked on, and let the orchestrator decide and log the rest. Keeping review policy in config, not in daemon code, is what made his replay evals cheap.

## 2. Eight agents with a message board did no better than eight working alone

**[Multi Agent Coordination Bench](https://benchmarks.bespokelabs.ai/multi-agent-coordination-bench/)** · Shreyas Pimpalgaonkar, Kartik Sharma, Carlos Bonilla and colleagues, Bespoke Labs · Bespoke Labs, 2026

MAC Bench is twelve four-hour tasks built so that each agent's work depends on its teammates'. In seven of them the team builds software in one shared repository. Opus 5.5, GPT-6.1 Sol and Gemini 4 Argon each ran in its developer's harness three ways: one agent, eight agents with a message board, and eight agents with no way to message. Averaged over the three models, a team of eight scored 0.42 and one agent 0.35. Eight agents without messaging added 0.087 over one agent. Allowing messages changed the score by -0.018, which the authors report as not statistically significant. Opus 5.5 was the exception: as a team of eight it reached 0.53, against 0.41 for GPT-6.1 Sol, though the two score about the same alone.

A team of eight cost 8.8 times as much as one agent at list prices and reached the single agent's final score a median 2.2 times sooner. In teams of two, 22% of tool calls already went to the message board. A teammate changed course within five minutes on 34% of 1,087 sampled findings. Duplicated work showed up in 64% of 139 team runs without damaging results; most damage came from messages nobody acted on and from parts that did not fit together. Claude models assigned the failure labels and checked them against the logs, and the harness has not been released, so outside readers cannot rerun the tasks yet.

**Why it matters:** Adding parallel agents to one task buys speed well below its cost, and on average the gain comes from extra workers, not from their messages. The failure tally points to interface ownership and acting on a teammate's findings as the places to add structure.
