Simon Willison released llm 0.35 yesterday, adding support for OpenAI’s GPT-6 Astra model to his command-line tool for working with large language models. If you’re not using llm yet, you should be. It’s a Python CLI that gives you instant access to pretty much every major LLM from a terminal prompt.

The new release adds the gpt-6-astra model identifier, which means you can now hit OpenAI’s latest model without switching contexts from your shell. Willison’s been maintaining llm as an open-source project that does one thing well: it abstracts away the annoying parts of working with different LLM APIs. You install it once, add your API keys, and then you’ve got a consistent interface whether you’re using GPT, Claude, Gemini, or local models.

This matters because GPT-6 Astra is OpenAI’s newest release, and having day-one CLI support means developers can script against it immediately. No waiting for official SDK updates, no writing your own API wrappers. Just llm -m gpt-6-astra "your prompt here" and you’re done.

The agent testing nobody asked for but everyone needed

While Willison was shipping practical tools, someone else did the unglamorous work of actually testing agent frameworks. A developer ran the same Three.js task across 10 different model and harness combinations to see which ones could actually complete it.

The results are worth reading if you’re evaluating agentic coding tools. Not all models perform equally in different harnesses, and not all harnesses get the same results from the same model. This is the kind of systematic testing that cuts through marketing claims.

Dan Luu also published a piece examining how well agents use test and verification techniques. The short version: not as well as you’d hope. Agents are getting better at writing code, but they’re still not great at the boring, essential work of verifying that code actually works. If you’re relying on agentic tools for production code, this is required reading.

Why OpenAI cares about defensive AI

Jakub Pachocki, OpenAI’s chief scientist, made an interesting argument about why the company needs to keep training smarter models quickly. The quote that stood out: “The strongest argument I see for continuing to train much smarter models quickly is the need to build defensive systems against the dangers posed by other AI.”

He’s talking about using powerful AI to secure infrastructure, protect against rogue agents in real time, and invent new protective measures. It’s a pragmatic take on the AI safety debate. Instead of slowing down development, OpenAI’s position is that you need more capable AI to defend against less aligned AI.

Whether you buy that argument or not, it’s worth understanding. This is how OpenAI is framing its deployment strategy going forward.

A trading framework nobody asked for

TauricResearch released a multi-agent LLM framework for financial trading on GitHub. It’s exactly what it sounds like: multiple AI agents coordinating to make trading decisions.

Should you use this? Probably not for actual trading. Should you look at the code if you’re building multi-agent systems? Maybe. It’s an interesting case study in agent coordination, even if the use case makes me nervous.

The framework is open source, so you can see how they’re orchestrating multiple models and handling the coordination logic. Just don’t blame me if you deploy it to production and lose money.

The bottom line

The practical stuff matters more than the hype. Willison’s llm tool keeps getting better because it solves a real problem: switching between LLM APIs is annoying, and having one CLI to rule them all saves time. The agent testing shows which tools actually work, not which ones have the best landing pages.

If you write code and you’re not using llm yet, install it. If you’re betting on agentic coding, read the test results. And if you’re building multi-agent systems, study the implementations even if you’d never use them for their stated purpose.

That’s what shipped this week that actually matters.