---
title: 'In the News: September 16, 2026'
description: 'OpenAI engineers detail the agentic pipeline running its production code; Anthropic explains how AI-driven CI growth broke its test system three times.'
canonical_url: 'https://darkfactory.dev/news/2026-09-16-morning'
markdown_url: 'https://darkfactory.dev/news/2026-09-16-morning.md'
collection: news
date_published: '2026-09-16T07:45:00-04:00'
date_modified: '2026-09-16T07:45:00-04:00'
---

# In the News: September 16, 2026


OpenAI's agentic pipeline now writes, reviews and deploys the company's production code, while an incident-response agent proposes mitigations without executing them. Anthropic's AI-driven growth broke its CI test-selection service three times in six months before the company completed a full redesign.

## 1. OpenAI's engineering leaders detail the pipeline now running its production code

**[Inside OpenAI's agentic software factory](https://newsletter.pragmaticengineer.com/p/openai-software-factory)** · Gergely Orosz, The Pragmatic Engineer · September 15, 2026

Gergely Orosz interviewed seven OpenAI engineers and engineering leaders, including Venkat Venkataramani, VP of Engineering for Applied Infra, about the pipeline that now runs the company's software. In Orosz's account, a human still defines the desired outcome, but Codex then gathers context from git, Slack, Notion, Databricks and Datadog, writes the change, runs it through CI and a new performance harness, sends it to review agents configured as domain specialists, and hands the approved change to a deployment agent that builds its own monitoring dashboard and watches the rollout. Venkataramani said pull request volume per engineer is "growing like a hockey stick," and that OpenAI is seeing "roughly a 10x increase in load on some systems" in about six months, growth he said would take two to three years at most companies. Production incidents go to Sevbot, an internal agent that gathers context and proposes mitigations in Slack but, in the piece's words, "never executes any." The published portion of the article covers this pipeline in detail; later sections on tooling changes, infrastructure scaling and how the engineering job itself is changing sit behind a paywall this edition did not access.

**Why it matters:** Orosz's account describes a coding-agent pipeline running a frontier lab's production software. It divides review among domain-specific agents, gives a deployment agent responsibility for rollout instrumentation, and prevents the incident-response agent from taking action. Teams assembling a harness can compare those boundaries with their own decisions about automation and human approval.

## 2. Anthropic's CI system broke three times before agentic coding forced a redesign

**[Agentic coding is straining CI. Here's how we scaled test impact analysis at Anthropic](https://claude.com/blog/agentic-coding-is-straining-ci-heres-how-we-scaled-test-impact-analysis-at-anthropic)** · Sachin Malhotra, Anthropic · claude.com/blog, September 14, 2026

Malhotra describes how Anthropic's CI job volume grew 25x over six months, driven by a codebase where the number of tests grew 10x while headcount barely moved. Engineers there now ship 8x as much code per quarter as they did between 2021 and 2025, and Claude authors 80% of that code. The service that decides which tests run on each pull request broke three times. Doubling the machine's core count bought time, though Malhotra writes he knew from the start it would be "fleeting." Sharding the service by code package, with Claude generating the split, bought 29 days. Daily restarts bought less than a day and quietly let some test results go unrecorded, so the service ended up selecting tests from stale data. The team then rebuilt the service as stateless and horizontally scalable, backed by a journal, a redesign that took three weeks for one engineer; a year earlier, Malhotra writes, the same project would have taken closer to a quarter.

**Why it matters:** Anthropic's growth rates show how code output can outpace the test infrastructure around it. Malhotra recommends planning for 10 to 20x the expected load and building stateless services that can scale horizontally.

## Also this cycle

- **[Adopting the software factory model: crawl, walk, run](https://x.com/zachlloydtweets/status/2099941244063432720)** · Zach Lloyd, Warp · X Article, September 15, 2026. Warp's CEO lays out a three-stage adoption path for agent-driven development and says the company's internal "walk factory" automates about 75% of changes to warp.dev. The piece closes with a pitch for Warp's own paid factory product, so read the adoption framework as a vendor's account of its own tooling.
- **[The Anatomy of Harness Engineering: How to Evaluate, Iterate, and Guard AI Coding Agents](https://x.com/GoogleCloudTech/status/2099946653134229721)** · Google Cloud Tech · X Article, September 15, 2026. Proposes "behavioral evaluation," discrete checks such as whether an agent asks a clarifying question or runs a validator before declaring a task done, as a complement to end-to-end benchmarks. The worked example runs specifically on Google's own Antigravity agent SDK.