---
title: 'In the News: August 27, 2026'
description: 'A five-person team banned human-written pull requests and published the numbers behind it: 295 issues, 217 shipped, in 30 days.'
canonical_url: 'https://darkfactory.dev/news/2026-08-27-morning'
markdown_url: 'https://darkfactory.dev/news/2026-08-27-morning.md'
collection: news
date_published: '2026-08-27T07:40:00-04:00'
date_modified: '2026-08-27T07:40:00-04:00'
---

# In the News: August 27, 2026


A five-person software team banned human-written code entirely and published the operating numbers behind that policy.

## 1. A five-person company banned human-written pull requests and published the numbers

**[Humans Architect The System, AI Writes The Code](https://tessl.io/blog/humans-architect-the-system-ai-writes-the-code)** · Paul Stack, System Initiative spinout · Tessl blog, Aug 19, 2026

In January, Paul Stack's team threw away a Rust codebase they had spent years building and rebuilt their process around agents from the start. The new rule: agents write every line of code, and a human-written pull request does not get merged, even an open source patch from an outside contributor. Work moves through a bounded state machine. One agent proposes a plan, a second agent attacks it for up to five rounds looking for security holes and missing requirements, and disagreements that survive go to a human arbiter. A file called CLAUDE.md acts as what Stack calls an executable contract, listing the constraints the agents must follow, and any failure that trips up a session is meant to become a permanent rule rather than a one-off fix. Over the previous 30 days the team opened 295 issues, shipped 217, and closed 81 as duplicates or work they weren't doing; median time from triage to ship was 1.6 hours. Five people spend roughly $3,000 a month on the workflow, most of it CI review. "Our rule is blunt: agents write every line of code," Stack writes. "If someone opens a pull request written by a human, we do not merge it."

**Why it matters:** Stack ties trust to five separate merge checks, including adversarial review and user-acceptance testing, rather than to the model doing the writing. A smaller team could adopt those gates one at a time, starting with a single constraint, without copying the whole policy.

## 2. Why an agent's answer can look right and still be wrong

**[An operating model for enterprise AI agent reliability](https://www.thoughtworks.com/en-us/insights/blog/generative-ai/operating-model-enterprise-ai-agent-reliability)** · Arun Srinivasan and Zichuan Xiong, Thoughtworks · Aug 14, 2026

A healthcare analytics agent was asked for Medicare Advantage readmission rates tied to a specific diagnosis and date range. It returned a plausible number that was wrong three separate ways: it mixed in the wrong plan type, used an outdated diagnosis code grouping, and misread the date range. "Our end-to-end evaluation missed all three, because the query ran and the output looked plausible," Srinivasan and Xiong write. Their proposed fix is a reliability ladder: six governance layers, terminology, routing, agent intent, semantic context, execution, and result, each owned by a different team and each backed by what they call a truth contract stating what must hold true, how it's measured, and what happens when it fails. A failure taxonomy assigns each kind of violation a code before it ever happens, so a broken contract routes automatically to whoever owns the fix.

**Why it matters:** "A final answer can be correct for the wrong reasons, or wrong even when every stage reports success," the authors write, which names the failure mode more precisely than most reliability write-ups manage. Testing only the final output, which is how most teams currently evaluate agents, will not catch this class of error.

---

## Also this cycle

- **[A Jagged Frontier: Evaluating Robustness of Code Agents to Semantics-Preserving Transformations](https://arxiv.org/abs/2608.18389)** · Mahmud, Gupta, Chaudhary, Enis, Mangal, Singh and Pasareanu (Colorado State, Microsoft, UIUC, Carnegie Mellon) · arXiv, Aug 18, 2026. Rewriting a codebase in ways that preserve its behavior (renaming variables, adding dead code) dropped agent resolve rates by up to 6.7 percentage points in the worst configurations, with no single model's robustness ranking holding across two different agent scaffolds. Which harness wraps the model may matter as much as which model it is.