---
title: 'In the News: August 15, 2026, Extra'
description: 'A solo entrant published the full agent harness behind a 232x kernel speedup, including the rule that stopped his agent getting stuck.'
canonical_url: 'https://darkfactory.dev/news/2026-08-15-extra'
markdown_url: 'https://darkfactory.dev/news/2026-08-15-extra.md'
collection: news
date_published: '2026-08-15T10:25:00-04:00'
date_modified: '2026-08-15T10:25:00-04:00'
---

# In the News: August 15, 2026, Extra


A contest write-up publishes its working harness and `AGENTS.md` in full, including the
change its author says got his agent out of a days-long rut. Debian has also opened a
two-week ballot on whether agent-assisted contributions are allowed at all.

## 1. A solo entrant published the harness behind a 232x kernel speedup, including the rule that unstuck it

**[Auto-research with codex: How I achieved a 232x Faster Kernel](https://sankalp.bearblog.dev/autoresearch/)** · sankalp (@dejavucoder), writing from his own contest entry · Personal blog, July 8, 2026

GPU Mode ran a contest to implement batched compact-Householder QR factorization on
B200 hardware. sankalp placed 12th of 183, taking the reference `torch.geqrf` path from
roughly 419,000 microseconds to a final tracked 1,805, and made over 1,500 leaderboard
submissions across 14 days doing it. He says he had known GPU kernel basics for about a
year and has not worked in the domain professionally.

The harness is printed in full, including the mistake it took him days to see:

> "For a long time, I did a dumb thing. I kept a single best candidate against which new
> candidates were tested. If the agent tries a new structural idea or a significantly big
> change, then it's highly likely that it would score less than our current best
> submission. However, after a few iterations, that change may outperform our best
> candidate. With this observation, I introduced some instructions to maintain a beam of
> 3-5 candidates."

A single-incumbent comparison kills every structural idea on its first measurement, which
is exactly when a structural idea is worst. The published `AGENTS.md` encodes the fix as a
standing rule: *"Do not optimize as a single-incumbent hill climb... Do not declare a
family dead after isolated singletons fail."* It requires at least three live beams, one
near the current best, one near-miss, one high-risk from profiling, each recording its
parent candidate, hypothesis and exact changed functions.

The same file carries an evidence-class rule most harnesses do not: *"Treat a timeout as
inconclusive, not as a correctness/performance rejection... Only completed pass/fail/timing
output is evidence."* Timeouts get counted as failures by default in most loops, which
silently retires whichever candidates are slowest to compile.

He states a criterion for whether a problem is loopable: an agent-usable CLI that can test,
benchmark **and
submit**, plus a checker that returns shape-wise feedback rather than pass or fail. Oversight
does not stop the run. Codex's `/goal` runs to a quantitative
objective, and one of his goals ran for over a day; `/btw` and `/side` open a side thread
carrying the main context, so a running loop can be questioned without pausing it. Pressing
escape, he notes, pauses the goal.

The article's own caveat matters. He attributes the escape from a plateau to the beam
change and to human steering, but nothing here is a controlled comparison: this is one
entrant's account of what he changed and what happened next. The externally scored parts
are the placement and the timings, which the contest checker produced.

Commenters supplied the boundary the article does not. `shken` wrote:

> "Every step here has an oracle: wall-clock, the profile, pass or fail from the verifier.
> I had an agent-built app audited task by task, 10 came back done and 7 worked, and the
> three misses were the ones needing a credential or a setting on someone else's dashboard.
> Nothing in the loop could tell the agent it had failed, so it said done and moved on."

`Almondsetat` reports running the same benchmark-profile-verify-research-improve loop on a
video codec that ships a bitstream verifier, and getting SSE and AVX implementations that
nearly doubled single-core performance in a couple of hours. Both are practitioner claims,
not measured results. The thread stood at 117 points and 38 comments at three hours old,
read at 10:13 EDT today.

**Why it matters:** the loop only pays where every step has a machine-checkable oracle, and
inside that boundary the harness rules are cheap and transferable. The beam-of-candidates
rule and the timeout-is-inconclusive rule can go into your own agent instructions this
afternoon, and neither depends on the domain being GPU kernels.

## 2. Debian opened a two-week ballot on whether agent-assisted contributions are allowed

**[General Resolution: LLM usage in Debian](https://www.debian.org/vote/2026/vote_002)** · Debian Project Secretary · debian.org, voting opened August 15, 2026

Voting runs from 00:00 UTC today to 23:59 UTC on August 28, after a discussion period that
opened July 23 and was extended. Eight options are on the ballot, drafted independently by
eight different proposers, which is why they do not read as variants of each other.

The spread is wide. Choice 1 amends the Social Contract to forbid contributions written
with LLM assistance outright. Choice 2 permits them under six conditions, including
disclosure via a Git trailer such as `Generated-By:` or `Assisted-By:`. Choice 5 declines
to endorse or prohibit and encourages but does not require disclosure. Choice 8 opposes
LLM use principally on climate grounds while conceding it "is just a position statement."

The majority requirement is asymmetric:
Choice 1 needs a 3:1 supermajority because it amends the Social Contract, and every other
option needs a simple majority. None of the options restrict upstream projects
Debian packages. Choice 7 states that it "does not cover indirect contributions, i.e. those
originating from upstream works," so a kernel or toolchain that accepts agent-assisted
patches is not in scope.

Choice 7 draws the line that will matter most to people who already work this way. It
disallows generative AI output as a direct contribution while explicitly not restricting
contributors from consuming such output when contributing: use the agent to explore,
research and critique, but the text you submit is yours. Its acknowledgements name GCC's AI
policy and rust-lang's LLM usage policy as its inspirations.

**Why it matters:** Debian is the first large distribution to put the question to a binding vote with the alternatives
written out rather than argued about. If you contribute to
Debian and use agents, the outcome decides whether you disclose, refrain, or carry on. The
disclosure mechanics being proposed, a commit trailer and an accountability clause naming
the submitter, are the shape other projects will copy either way, and the asymmetric
majority bar means the outright ban is the hardest of the eight to pass.