---
title: 'In the News: November 15, 2026 (Midday)'
description: 'Justin Searls says his agents ship 20 to 40 issues a day while he answers their questions from a phone inbox, and he published the build spec.'
canonical_url: 'https://darkfactory.dev/news/2026-11-15-midday'
markdown_url: 'https://darkfactory.dev/news/2026-11-15-midday.md'
collection: news
date_published: '2026-11-15T13:31:00-05:00'
date_modified: '2026-11-15T13:31:00-05:00'
---

# In the News: November 15, 2026 (Midday)


Justin Searls has published the design of the loop that runs his solo iOS work: in-app bug reports go to agents, and only the questions agents cannot answer come back to him, as cards on his phone. He wrote it up as a spec meant to be handed to your own agent. The second item is a job description from Qodo's chief executive for the engineer who owns a company's software factory.

## 1. Searls hands his backlog to agents and keeps only the decisions

**[I'm just a meat proxy](https://justin.searls.co/posts/im-just-a-meat-proxy/)** · Justin Searls, co-founder of Test Double · justin.searls.co, November 11, 2026

Searls writes: "My agents are currently shipping 20 to 40 issues a day across the dozen or so iOS apps I'm building." Every app has a reporting button that files a GitHub issue with the last minute of screen recording and his spoken narration, logs, and a deep link to the screen he was on. Scheduled jobs on two Macs give each issue a fresh agent; bugs are reproduced first and ship with before and after recordings, and a physical iPhone and iPad stay attached for agents to test on. In his account, "6 to 10 issues get a human label slapped on them" each day because an agent is stuck, and those become cards in an app he calls MeatProxy. A daily retro agent reads the previous 24 hours and proposes changes; he says one retro found and repaired the watchdog that should have killed a set of stuck agents.

He does not release the app. He posts a long build spec instead, and it is specific about where autonomy stops. "Code decides, models work": scheduling, eligibility, retries and budget gates are plain scripts. Review comes from "a fresh session that did not write the code, ideally a different model, never a self-review." Two unfinished attempts at an issue send it to a human. On permissions he is blunt: headless agents run with prompts bypassed, and "Bypass is what makes the loop work and it means an issue's text can run commands," so the spec pairs that with a trusted-author rule, a scoped bot account and a Mac holding no production credentials. The throughput figures are his own and unaudited, and he concedes he is "almost certainly accumulating a mountain of technical debt."

**Why it matters:** This is a working single-developer factory with its control points written down: what a script decides, what a model does, when a human is pulled in, and the security cost of running unattended. The spec is directly reusable as a checklist even if you never build the phone app.

## 2. Qodo's CEO drafts the job description for a factory engineer

**[Software Factory Engineer: Job Description](https://x.com/itamar_mar/status/2118356299612033305)** · Itamar Friedman, CEO and co-founder of Qodo · X article, November 5, 2026

Friedman, whose company sells code review for agents, describes a role responsible for moving a team to what he calls an "L4 Autonomous Software Factory," one that can execute, verify and ship defined classes of change and escalate when stakes or uncertainty are too high. "Your primary engineering artifact is the Software Factory itself." The verification section is the most concrete: "The author should not self-approve," and the authoring agent must not be able to redefine its own acceptance criteria. He lists factory metrics: cost per accepted change, time from intent to validated production outcome, escaped defects, verification quality, human intervention and recovery time. He adds a caution: "Measure human verification effort and cognitive burden alongside intervention frequency: fewer interventions can still conceal longer, harder investigations." It is a proposal and a hiring template, with no deployment data behind it.

**Why it matters:** It gives teams a vocabulary and a metric set for treating the factory as a production system with an owner. The intervention-count warning is the line to take: a factory can look more autonomous while costing reviewers more.
