---
title: 'In the News: August 15, 2026, Evening'
description: 'A BDD maintainer shipped a release written mostly by an agent and named the rule that kept its test suite honest: never edit code and tests in the same step.'
canonical_url: 'https://darkfactory.dev/news/2026-08-15-evening'
markdown_url: 'https://darkfactory.dev/news/2026-08-15-evening.md'
collection: news
date_published: '2026-08-15T20:45:00-04:00'
date_modified: '2026-08-15T20:45:00-04:00'
---

# In the News: August 15, 2026, Evening


Do not let an agent change production code and the tests for that code in the same step. Stephen Cresswell explains why in his write-up of the Yadda 3.0.0 release, which he had Claude Code do most of the work on. A June 2025 essay by Venkatesh Rao reframes the "cognitive debt" finding as a management problem and resurfaced this week in a live discussion of the same idea.

## 1. A maintainer had an agent modernise his library, and phase-separating the test edits is what kept the suite as evidence

**[Yadda 3.0.0: BDD in the Age of AI Agents](http://www.stephen-cresswell.com/2026/08/15/Yadda-3.0.0-BDD-in-the-Age-of-AI-Agents.html)** · Stephen Cresswell, maintainer of Yadda (acuminous) · Signal Over Noise, 15 August 2026

Cresswell published [Yadda 3.0.0](https://www.npmjs.com/package/yadda), a modernisation of his JavaScript BDD library: Node only, `node:test`, ES6 source, TypeScript definitions, and the removal of CasperJS, PhantomJS, Bower and Component. He reports that he did it with Claude Code on Opus 4.8, working from a phased epic that the agent itself wrote, and that from starting the work to publishing the package took roughly a day of elapsed time while he was doing other things in parallel. That figure and his assessment that the agent "made remarkably few mistakes" are his own account of his own project, not a measurement.

He imposed one reusable constraint:

> "I also deliberately avoided asking Claude to modify production code and the corresponding tests in the same step. If an agent changes both simultaneously, a green test suite becomes weaker evidence because it is free to change the definition of 'correct' at the same time as the implementation."

The rest of the piece argues that executable specifications are worth more now than when BDD was invented for human readers. A wiki, he writes, "cannot, by itself, tell you whether the system actually does it. An executable specification can." Yadda 3.1.0 added GitHub-flavoured Markdown as a spec format so the specification can sit beside the wiki and the issues that agents already read. He describes the resulting artifact as a contract rather than documentation: "An implementation agent can use it to understand the required behaviour. A testing agent can use it to determine what needs validating. A reviewing agent can use it to challenge an implementation. CI can continuously verify it."

He also names his own ceiling, and it is not the model. Running several agents in parallel via worktrees and session managers, he found he could keep three tasks moving comfortably, sometimes four or five, before losing track of which decision belonged to which agent: "the model is not overloaded and the machine is not overloaded. The bottleneck is the human coordinating the work."

The [Hacker News thread](https://news.ycombinator.com/item?id=49310495) stood at 54 points and 27 comments about ten hours after submission, read at 20:20 EDT. It is not uniformly agreeing. One commenter asks why natural language is the right abstraction layer at all when every other abstraction in a codebase is written in the programming language; another argues BDD did not survive its first decade. A third reports having run transcript to specification to issues for close to a year.

**Why it matters:** The phase-separation rule is the cheapest verification fix in this edition and it does not require adopting BDD. A test suite is only external evidence while the agent cannot edit it in the same breath as the code it is meant to constrain. Everything else here, the executable specification as a shared contract across implementation, review and CI, is a bet that the expensive half of BDD was writing the specification and that the cost has now moved.

## 2. The "cognitive debt" result has prior art that reads it as role confusion, not neurology

**[Prompting is Managing](https://contraptions.venkateshrao.com/p/prompting-is-managing)** · Venkatesh Rao · Contraptions, 19 June 2025

Rao's essay responds to the *Your brain on ChatGPT* preprint, the EEG study in which undergraduates leaning on GPT-4o showed weaker alpha-beta coupling, produced similar prose, and later could not quote their own sentences. He accepts the data and rejects the interpretation. His claim, stated as a principle:

> "For today's text generators, the cognitive effects of prompting an LLM are empirically indistinguishable from supervising a junior human."

On that reading the flat neural traces are the signature of supervisory control, the delegate, monitor, integrate, ship loop that professors run with grad students and editors with freelancers, and the students failed because "freshman comp doesn't teach management." He calls them accidental managers, and prescribes delegation protocols, quality gates and exception handling rather than abstinence. He also relocates the risk he does take seriously, away from the brain: an "index-fund pull toward the linguistic mean, cheap, reliable, originality-suppressing."

The essay is self-disclosed AI-assisted work: it publishes its own recipe and links the model transcript that produced it, and at least one commenter disputes whose argument it is on that basis. It is also fourteen months old. It is here because the argument now running about whether agents dull the people supervising them is being had largely without it.

**Why it matters:** If prompting is supervision, then the things worth measuring are the supervisor's, not the model's: how fast errors are caught, what the quality gate actually gates, what the exception path is when the agent returns something plausible and wrong. That is the same list a harness has to implement, and it suggests the skill gap in agent-assisted work is a management gap with a long literature rather than a new deficit.