---
title: 'In the News: November 22, 2026 (Extra 2)'
description: 'Anil Madhavapeddy says better harnesses alone will not fix software, and argues for formal specs plus deliberately varied implementations.'
canonical_url: 'https://darkfactory.dev/news/2026-11-22-extra-2'
markdown_url: 'https://darkfactory.dev/news/2026-11-22-extra-2.md'
collection: news
date_published: '2026-11-22T16:01:00-05:00'
date_modified: '2026-11-22T16:01:00-05:00'
---

# In the News: November 22, 2026 (Extra 2)


A co-author of the original unikernels paper wrote in April that he could not see every layer of software ever being formally specified. He now says coding agents have changed his mind, and that better harnesses and models will not rescue software quality without specs to derive code from.

## 1. Anil Madhavapeddy: harnesses alone will not fix software, so derive it from specs

**[Was software just always too complex for humans?](https://anil.recoil.org/notes/software-too-complex-for-humans)** · Anil Madhavapeddy, co-author of the 2013 unikernels paper · anil.recoil.org, November 22, 2026

Madhavapeddy quotes his own April post, which doubted that all layers of hardware and software would ever be formally specified, and then reverses it. Given the pace of frontier models, he now thinks it plausible to write formal specifications for existing services first, translate legacy code against them, and derive all new software from those specs. He argues that unikernels, with fewer interfaces, make the specification job smaller, and that fixes re-derived from a spec and machine-checked could take humans out of the security patch loop.

His second proposal is deliberate diversity. Because a good spec allows several valid implementations, agents could produce varied ones so that one bug does not hit every host at once, and he says that diversity "has to extend beyond the code to the models and harnesses too." His answer to his own title: "I don't think better harnesses and models alone will let software engineering catch up, as they just produce mountains of slop code." Humans, in his account, still decide what the specs contain. The note is an argument and reports no measurements.

**Why it matters:** Madhavapeddy puts the lasting value of an agent-built system in the spec and the checker, and treats a single shared model or harness as a monoculture risk. Both points give a team something to check now: what its verification loop checks against, and how many of its services came out of the same model.

---

## Also this cycle

- **[Agentics: Mushy Software](https://12gramsofcarbon.com/p/agentics-mushy-software)** · theahura, founder of Nori · 12 Grams of Carbon, November 18, 2026 · Describes how Nori runs its internal tools: each has one long-running "PM agent" that farms tasks out to subworkers, builds in an isolated git repo and VPC, deploys to AWS and tracks bugs live. He says the PM agent automatically tries to fix failures and retry; the post gives no figures on how often that works.
