Agentic Engineering Roundup: Seven Days with Fable, the Billing Carnage, Reading the Code

Here we go with a mid-week roundup.

20 mins each morning and 20 each evening with Fable, one technique for long horizon task shared for each day. A quick write-up Seven Days with Fable & Seven Practical Skills with ‘burner boxes’ being the most commented thing people found interesting.

I hand write all the copy, and Fable drives crazy amounts of automation of a quite locked down spare Macbook I have. Has notes on: goals and statements of work, an autonomous test loop, interactive specs instead of prototypes, soak testing, talking instead of typing, anchoring to idioms, and managing your blast radius.

Signalbox is the result and the write up is on my blog.

I wrote up the seven techniques that made the long unattended runs work -

Full write-up here:

The backdrop to my week of actually using Fable was everyone else arguing about paying for it. Last issue was the outrage; this week Anthropic conceded: Fable 5 is now permanently included in Max plans (Simon Willison’s take). The catch was spotted within minutes on the Reddit megathread: bonus usage ends the same day, which nets out to roughly a third less usage overall.

Meanwhile The Kimi K3 Moment is the essay of the week on why practitioners are moving, and demand ran hot enough that Moonshot stopped taking new subscriptions. Personally I’m forced to always pair Fable with GPT-5.6 Sol so Sol takes all security work (Fable drops to Opus otherwise) but I actually prefer pairing that way. And it is somewhat scary how quickly Fable was able to run my mouse, the iOS emulator, take screenshots, automate the MacOS menu-bar. Blast radius more important than ever!

Mitchell Hashimoto was asked how he handles AI-generated code and answered in four words: “I read the code.”. This was the best four words I read this week. Also Reverse-engineering is cheap now - agents make reading other people’s code cheap too.

Most funny is how larger publications don’t use “Hugging Face” cause I guess it doesn’t translate well to their headlines - instead:

• CNN: “an OpenAI test model escaped and broke into a real company’s servers”

• The Guardian: “AI agent went rogue and hacked a prominent startup by itself”

• Euronews: “OpenAI models autonomously hacked another AI company”

• NPR: “a digital library of AI technology”

• BBC: “a technology start-up”, then “one of the world’s largest open-source hubs for sharing AI models”

• And the one that just said it, CNBC: “OpenAI cyber models broke out of training environment to hack Hugging Face”

A counterweight to all the billing noise: Fable produced a counterexample to the Jacobian Conjecture, and Terence Tao’s follow-up made it the biggest general-AI story of the week.