WEEK OF AUGUST 8–14, 2026

Five things happened this week that are worth your ten minutes: a 150-million-parameter model quietly out-argued the "just add more compute" crowd, DeepSeek's agent harness pulled in more stars than anything else on GitHub, and Anthropic started tattooing invisible marks into everything Claude writes. Here's what actually mattered, sorted the way we always sort it.

— three papers worth the click

A 150M-parameter model just posted a new cost-accuracy frontier on ARC-AGI-1 by looping reasoning through a recurrent latent space instead of scaling up parameters. Worth reading if you've been assuming reasoning gains only come from bigger checkpoints — this argues otherwise, and did it at a fraction of the size.

→ arxiv.org/abs/2608.09888 · Hugging Face

An agent that reviews and rewrites its own tooling, prompts, and core code — and posted new records doing it: 86.74% on Terminal-Bench 2.1, 90.69% on OSWorld-Verified. The more interesting detail is buried in the appendix: "Hope," its longest-running deployment, has been evolving itself for 161 straight days under human oversight.

→ arxiv.org/abs/2608.08311 · Hugging Face

A lightweight skill for coding assistants that separates planning from writing and forces every claim through an evidence check before it's allowed to stand — a direct answer to the flood of fabricated citations showing up in AI-assisted papers. Picked up 270 upvotes in two days, which tells you the problem it's solving is a real one.

→ arxiv.org/abs/2608.11924 · Hugging Face

— the repos everyone starred this week

DeepSeek's own harness for running its models gained more stars than any other repo on GitHub this week — over 32,000 of them. Whatever you think of DeepSeek as a company, the community reaction says something real about how badly people want an open alternative to Claude Code and Codex.

→ github.com/deepseek-ai/deepseek-harness

Twenty-nine editorial diagram types built specifically for Claude Code — self-contained SVG, no Mermaid, no drop shadows. It's a small idea, but it's the kind of small idea that a lot of people apparently needed: 10,800 stars in a single week.

→ github.com/cathrynlavery/diagram-design

A full roster of specialized AI personas — frontend, community management, even a dedicated "reality checker" — meant to run like a tiny agency out of one repo. Half novelty, half genuinely useful scaffolding if you're a solo builder trying to look like a team.

→ github.com/msitarzewski/agency-agents

— what the front page argued about

Landed at #1 on the HN front page the same day it topped GitHub's weekly trending list. The thread is where the real argument is happening — open ambitions versus the actual rough edges people are running into.

→ Hacker News thread

Florian Herrengt's post on how coding agents let a mediocre engineer ship bad decisions ten times faster — and how untangling one bad AI-assisted database migration eats more time than five new pull requests create. 314 points, 364 comments, and a lot of senior engineers nodding along uncomfortably.

→ blog.florianherrengt.com · HN thread

A reminder that "AI agent" user-agent strings are now valuable enough to fake outright, and that trusting a bot by its name alone is getting riskier by the week.

→ knownagents.com

— worth watching, not just reading about

Codex, Claude Code, Qwen3.8-Max, and Kimi K3 run the same three tasks on the same machine, every result graded on camera. A rare head-to-head with actual receipts instead of vibes — quality tied 12/12 across the board, so the real story is speed and cost.

→ Watch on YouTube

A fast, no-nonsense breakdown of xAI's Grok 4.6 launch and why it's landing right next to GPT-5.6 Sol on the Artificial Analysis Intelligence Index at less than half the price per token.

→ Watch on YouTube

— what Anthropic, xAI, and OpenAI shipped or said

New Claude models launched on or after August 2 now embed an invisible, machine-readable mark into generated text, plus signed C2PA metadata on image files — timed to Article 50 of the EU AI Act taking effect, but applied worldwide rather than region by region. Heavy edits or format conversions can still strip it, and Anthropic has said a public detection API is coming.

→ TechCrunch

Not a new base model — xAI kept the Grok 4.5 foundation and bought the gains through a longer supplemental training run and expanded reinforcement learning. The Intelligence Index score jumped five points to 61, tying GPT-5.6 Sol, while finishing long agentic jobs in roughly half the turns Opus 5 needs. Pricing stayed flat at $2 / $6 per million tokens.

→ x.ai/news/grok-4-6

ChatGPT Desktop's Codex environment is now available on Linux, closing a gap that's been a running complaint from developers who do the bulk of their agentic coding work outside macOS and Windows.

→ openai.com/codex

That's the week. See you next Tuesday.