---
title: 'In the News: October 13, 2026 (Extra 3)'
description: 'A public math repo with hundreds of open pull requests ships a written protocol for donated agent sessions: task allocator, overlap checker, advisory claims and maintainer-only merges.'
canonical_url: 'https://darkfactory.dev/news/2026-10-13-extra-3'
markdown_url: 'https://darkfactory.dev/news/2026-10-13-extra-3.md'
collection: news
date_published: '2026-10-13T23:05:00-04:00'
date_modified: '2026-10-13T23:05:00-04:00'
---

# In the News: October 13, 2026 (Extra 3)


One public repository now documents how hundreds of donated agent sessions should pick, claim, check and hand off work without a human assigning each one. It is a math project, not a software product, but the coordination rules are the useful part. A second item covers a game developer's list of what to build so an agent can iterate on a game quickly.

## 1. A math repo with hundreds of open pull requests publishes the operating rules it gives donated agents

**[integer-mult-bounds](https://github.com/CrocSwap/integer-mult-bounds)** · Douglas Colkitt (maintainer) and community contributors · GitHub, README and autoresearch.md read October 13, 2026

The repository collects community attempts to improve a conditional bound for integer multiplication, building on an OpenAI manuscript. Its README says Douglas Colkitt maintains the project with OpenAI Codex assistance, and it invites donated agent sessions: anyone "asking an agent to contribute" is pointed to `autoresearch.md`. At the builder's October 13 read, the page listed 721 open pull requests; that live count changes rapidly. The current reviewed result is a conditional exponent saving of about 0.00167266, submitted as Rohan Arun's pull request #1094. The README calls that an exponent saving, "not measured runtime."

The protocol is the part worth reading. A donated session defaults to about 60 minutes. A script draws a random task from weighted tracks (bit mechanisms and complex mechanisms 20% each, adversarial audit 15%, independent verification 10%, and others), and a second script searches open PRs, discussions and notes for overlap. Agents post a claim on a discussion board with an expiry of roughly two hours; claims are advisory, not locks. The file says plainly that "Random choice is not an atomic work allocator: hundreds of agents still need the live discussion and PR overlap checks." Merging, closing other people's work and changing the selected result stay with the maintainer, and finishing a session with a finding or handoff and no PR is allowed. Each reviewed result ships with a make target that re-runs its checks. Rohan Arun, a credited contributor, posted on X that anyone could contribute by "copying this prompt into Codex/Claude."

Several limits apply. The README states that the work "is not full formal verification, independent human peer review, a worldwide priority claim, or a practical multiplication benchmark," and the bounds depend on an assumed framework. The README does not say what share of the pull requests were written by agents, and this edition did not check the contributor notices. A tracker page hosted by Super Powers AI reports 234 contributors; we did not verify that count.

**Why it matters:** If you plan to point many unsupervised agent sessions at one repository, this is a working example of the rules that become necessary: a task allocator, an overlap check before work starts, short expiring claims, a record of failed experiments, and a maintainer who alone decides what counts as accepted. The repository does not show that the protocol works well, only that its maintainers consider it necessary at this volume.

## 2. A game developer lists what to build so an agent can run a fast edit-and-check loop

**[Your Game Is A Harness](https://kylekukshtel.com/ai-harness-engineering-game-development-claude-code)** · Kyle Kukshtel, game developer · kylekukshtel.com, October 10, 2026 ([HN discussion](https://news.ycombinator.com/item?id=50072120))

Kukshtel, who says he has programmed games for nearly 15 years, wrote up a talk he gave at Indie Rodeo in Calgary on September 12 and states that the post is entirely human-written. His argument is that agents fail on games because much of a game is not code, so the fix is to make the game addressable. He lists concrete steps: one seeded source of randomness so any state can be reproduced; saving and loading from the start of the project; a single object type for everything in the game; serializable commands that an agent or a text file can replay against a running build; static data kept outside the code so values change without recompiling; and a simulation core in its own library that takes commands and returns state.

Two of his points are specific enough to try. He argues that handing a model before and after screenshots adds noise, and suggests the free odiff tool, which produces a diff image and a similarity figure ("85% similar") the agent can track from turn to turn. For performance he prefers samply, because it can dump profiles as text an agent can read. His line on the goal is "The tighter you can make the agentic loops, the better!"

He reports no measurements. The claims that screenshot pairs confuse models and that a full loop can run "in seconds" are his own assertions. His postscript says he is shipping a small project that used none of this, and that the approach is aimed at larger games.

**Why it matters:** For anyone building agent loops around software that is not mostly text, the post gives a short checklist of what to expose: reproducible state, a numeric progress signal, replayable commands and text-readable profiles. Treat it as one experienced developer's practice, not tested guidance.