Kurt Thorn

GitHub - tkellogg/open-strix Contribute to tkellogg/open-strix development by creating an account on GitHub.

open-strix: an opinionated agent with a small feature set, focused on stable & sustainable agents

uvx open-strix setup --home my-agent --github

github.com/tkellogg/ope...

An AI agent coding skeptic tries AI agent coding, in excessive detail No vagueposting here, just look at the Estimated Read Time.

New blog post up: the culmination of my past few months working with agents Opus 4.5 and beyond, and the *many* things I learned. Also, the discovery of an optimization trick with promise.

As a bonus: this post will make Rust engineers very mad. minimaxir.com/2026/02/ai-a...

Hyperion author Dan Simmons dies from stroke at 77 I went into Hyperion blind, decades ago, knowing almost nothing about it. I was never the same.

RIP. A visionary writer of sf and adjacent genres.

This makes me happy. Centralizing billing through AWS is nice

The entire Democratic strategy debate — "moderate" or move left? fight or compromise? — is based on a misreading of data. Voters don't think Dems are too progressive; they think Dems are weak. The actual math on what to do about this isn't even close!
New: www.gelliottmorris.com/p/new-poll-d...

I have a lot of fond memories of playing PVP Marathon in lab in college while waiting for X-ray diffraction experiments to run.

Improving AI Agent Performance with Domain-Specific Skills | PyMC Benchmark Study A controlled benchmark comparing base LLM performance to skill-augmented AI agents for Bayesian modeling. Domain skills significantly improve convergence, model specification, and reliability.

We gave Claude a 15-page PyMC best-practices cheat sheet.
Pass rates jumped from 60% → 93%.

Hardest tasks saw the biggest gains:
• Stochastic volatility: 0% → 67%
• Runtime: 19 min → <3 min
and more ..

👉Benchmark: dub.link/KZxxZfQ
👉 Get the skill: hub.decision.ai/skills/pymc-...

#BayesianModeling

WTF Happened in 2025? A collection of datapoints in or around 2025 which we may look back as a historical inflection point. Open source & curated in realtime by Latent Space.

presented without comment wtfhappened2025.com

There is a lot of concern about the lack of transparency and predictability of the FDA when it comes to getting new trials approved.

The FDA traditionally doesn’t do a lot of research- it reviews and approves clinical trial plans and then decides whether or not to approve drugs based on the results. That is definitely disrupted - see the recent denial of Moderna’s vaccine trial and then their reversal.

I think any historic documents from the FDA are probably fine. There’s a lot of useful info there. I’d only question stuff from the last year. (I work in drug discovery)

Claude code remote control is cool. Finishing up some work while I walk to an event at my kids’ school

my local park is full of hundreds of snow sculptures and someone has been adding museum labels

Another duo of agents dropping today:

El Agente Gráfico
📃 arxiv.org/abs/2602.17902
📽️ www.youtube.com/playlist?lis...

El Agente Sólido
📃 arxiv.org/abs/2602.17886

[1/7]

Man accidentally gains control of 7,000 robot vacuums Sammy Azdoufal just wanted to steer his DJI Romo with a gaming controller.

These kinds of "incidents" will only increase as AI chatbots make it easier for anyone to probe security flaws in everyday devices. Do we really need all these internet-connected gadgets at home?
#IoT #Cybersecurity #AI

OK, you can now see my D&D / sci-fi weapon generator at infinitearmory.fly.dev/weapons. Source code at github.com/kthorn/infin...

Claude in PowerPoint: A New Add-in for Building and Editing Slides Using AI Anthropic launched Claude in PowerPoint as an add-in to help professionals build and edit slides using conversational AI.

Anthropic launched Claude in PowerPoint as a Microsoft AppSource add-in (research preview) that embeds Claude into active PowerPoint to generate native editable slides, support three workflows, read slide masters, and warn of beta security and compliance limitations.

100%

Now expanded to do sci-fi weapons and mechs. I need to get auth and cost controls in place and then I’ll share it

Also, AI models are way more creative than I am - my son asked for a weapon with affinity for sloths and got back “Sloth’s Sanctuary Sword” that can protect a sloth within 60’.

AI coding is really fun. On Friday, my 8yo son asked for an AI generator for D&D style weapons - you give a verbal description and the AI generates a name, description, and stat block plus an image. Today I deployed a first draft at fly.io - all coding was done by Claude. Took me about 4hr hands on.

Forget wealth tax: Should state revamp Prop. 13? USF prof says it’s time to touch third rail.

We should actually be working on fixing the state's "original wealth tax" PROPERTY TAXES:
www.sfexaminer.com/news/politic...

The California Assembly Bill 2047 document, tilted at a jaunty angle over the California Republic flag, conveniently blocking the bear... which is probably also on a state watchlist for having claws capable of subtractive manufacturing.

California's New Bill Requires DOJ-Approved 3D Printers That Report on Themselves

Your 3D printer needs to be on a DOJ-approved list or it can't be sold in California. Disabling the surveillance software? That's a misdemeanor. 🐻

blog.adafruit.com/2026/02/19/c...

This is just going to be a bunch of musings, something I haven’t been able to get out of my head for a while now. Inspired by my many text thread and in-person conversations with “old timers” and veterans like myself who have watched a real paradigm shift in the drinks trade lately.

Opinion | The A.I. Disruption Is Actually Here, and It’s Not Terrible

NYT OpEd asked me to explain vibe coding to a general audience, and I took a swing. www.nytimes.com/2026/02/18/o...

Not ai not cgi - real robots

… I work with a lot of unmaintained academic packages, why do you ask?

This is a good idea. Yesterday Claude code monkey patched a bunch of imports in a dockerfile because the upstream package had incompatible dependencies. Another time I forked a repo to fix an issue and Claude promptly found a bug where the original code used min instead of max.

How System Prompts Define Agent Behavior System prompts matter far more than most assume. A given model sets the theoretical ceiling of an agent’s performance, but the system prompt determines whether this peak is reached.

Srihari Sriraman and I took a deep dive into coding agent system prompts to understand their structure, similarities, and differences. www.dbreunig.com/2026/02/10/s...

Harness engineering: leveraging Codex in an agent-first world By Ryan Lopopolo, Member of the Technical Staff

openai.com/index/harnes...

i have been mildly obsessed with the ideas in here for the past month or so.

incredibly validating, total victory: my suspicions are correct, actually. "This functions like garbage collection" is a sentence I have said out loud. craziness