---
title: 'In the News: September 22, 2026 (Evening)'
description: 'Ten agents on a message board produced a Lean-checked shortest-path bound. The proof package says the peer reviews were done by the agents themselves.'
canonical_url: 'https://darkfactory.dev/news/2026-09-22-evening'
markdown_url: 'https://darkfactory.dev/news/2026-09-22-evening.md'
collection: news
date_published: '2026-09-22T19:10:00-04:00'
date_modified: '2026-09-22T19:10:00-04:00'
---

# In the News: September 22, 2026 (Evening)


A Vals AI researcher put ten Claude Opus 5.5 agents on a shared message board, pointed them at an open algorithms problem, and refused to accept a result without a Lean proof. The formal check held. The proof package is also candid that the "peer review" step was the agents reviewing each other. Separately, Josh Rosen of ThruWire argues that a software factory's run history is what a team should be designing for.

## 1. Ten agents, a message board and a Lean gate produce a checked shortest-path bound

**[A Faster Shortest Path Algorithm](https://www.vals.ai/blogs/faster-shortest-path-algorithm)** · Geby Jaff, Vals AI · vals.ai, September 20, 2026

Jaff writes that he chose to "spawn 10 Claude Opus 5.5 agents at maximum effort and give them a simple message board." The agents started with assigned roles but could reorganize. They were told to record failed approaches so others would not repeat them, and they could not declare success without "a reproducible Lean build and two separate peer reviews." After about 15 hours and 733 messages they produced C-HD, an algorithm for exact single-source shortest paths on directed graphs with non-negative real weights. Its proved bound along one density profile is O(n log^(11/12) n), against O(n log n) for Dijkstra. Jaff states the limits himself: "This is not a measured speedup." The constants are enormous, and the result says nothing about denser graphs such as m = 10n.

The [proof package on GitHub](https://github.com/spicylemonade/c-hd-proof) is more careful than the post on one point. Its README reports that the full build passed (2,548 jobs) and that two kernel replays each checked 18,994 constants. It also says the two reviews "are reviews by the research agents, not outside peer review," that the second is a snapshot with reproduction fields still pending, and that novelty is a matter for external experts. We read the README; we did not rebuild the proof. The post is dated September 20 and names Claude Opus 5.5, which Anthropic announced on September 22. It does not say how the author had access. The Hacker News submission had 29 points and 7 comments at about 22:10 UTC.

**Why it matters:** The part that held up is the part no agent could argue with: the Lean kernel accepts the theorem or it does not. The review step, by the package's own account, was agents grading agents. If you run a multi-agent loop on work that matters, put a mechanical check at the finish line and do not let agent-to-agent review stand in for it.

## 2. Josh Rosen: a software factory's run history is the asset

**[Software Factories: It's All About the Data (Architecture)](https://x.com/JoshARosen/status/2102457068640711142)** · Josh Rosen, ThruWire · X Article, September 22, 2026

Rosen's thesis is that "software factories are data factories." He describes four data layers, each drawn from another vendor's product: OpenTelemetry trace export from Factory's Droid; scorers and LLM-as-a-judge run over historical traces at Warp; Warp Factory Benchmarks, which replay past runs against different models, harnesses, skills and prompts; and Port's Context Lake, which joins an agent run to its PR, its deployment and what happened in production. He then adds his own company's approach: durable checkpoints around the agent loop, where the model "has to show its work in a form the system can understand." The article reports no measurements. On X it had 67 likes, 104 bookmarks and 4,856 views at about 19:10 Eastern, five hours after posting.

**Why it matters:** Replay is the practical idea here. If past runs are stored with their starting state, a rewritten skill or a new model can be tested against real historical work before the factory depends on it. That only works for runs whose traces you keep, so the decision is about what to log today. The vendor tour is Rosen's summary of other companies' products, and the piece ends at the one his company sells.