Do the models you use have an AI bill of materials? | CIQ’s Gregory Kurtzer
Do the models you use have an AI bill of materials? | CIQ’s Gregory Kurtzer
Front row at AGNTCon, the possibility matrix, and is your bot too fat?
Most companies control the prompts they send to a model, but not what went into it. That’s the gap CIQ founder and CEO Gregory Kurtzer, who set up the community that built Rocky Linux, wants an AI bill of materials to close. This week on Dev Interrupted, he joins Andrew to make the case for OpenWALDO, a community-backed corpus of AI training data built on open weights, artifacts, licenses, data and origins. Greg explains why open weights alone don’t make a model open source, and why a model trained on copyleft code raises a licensing question nobody has litigated. They close the conversation on why the community side of open source matters more than ever now that AI lets more people contribute.
1. The Agentic AI Foundation grows its roster
I spent the end of last week in the audience at AGNTCon + MCPCon North America, and the AAIF saved its big news for the opening keynotes. Four new hosted projects took the foundation from three at its launch last December to ten. Anthropic contributed Agent Skills, OpenAI brought Open Responses, Hugging Face added OpenEnv for reinforcement learning environments, and Alibaba handed over OpenSandbox. Between capabilities, model interoperability, training and secure execution, the AAIF is going after the whole agent stack, and it was a lot of fun to watch it unfold from the seats. I also sat down with Angie Jones and Mazin Gilbert while I was there, so stay tuned for those interviews here on the Substack.
Read: Agentic AI Foundation Announces Four New Projects
2. Your AI skills have an expiration date
Not every AI skill comes packaged in a folder, and Charles Miller doesn’t expect the human kind, the knack for getting good results out of a model, to keep. When he was learning C in the mid-1990s, it was common knowledge that ++i ran faster than i++, even though optimizing compilers had already turned that into folklore. He expects today’s prompting tricks to go the same way, folded into products as default behavior within a few months, which leaves them useful for doing the job now and worth very little after. I’d add that the announcements in San Jose are one way that happens, since a trick becomes a format, the format becomes a standard, and the standard finds a home at a foundation where nobody has to think about it anymore.
Read: “AI Skills” are worthless in 2026
3. A possibility matrix for the edge of what you know
When an agent asks what you think, it’s asking you to lay out every possible answer, which is the one thing you can’t do at the edge of your knowledge. A possibility matrix does that part for you, with each row a defensible option and a Holds if column naming the premise that has to be true for it. You stop picking answers and start verifying conditions, with an unsure row, a reframe row and a closing prompt asking what row is missing. Dakota Fabro labeled 10 findings from a race-condition checker they’re building, then went back through them with an agent and a matrix for each, and five labels changed, every one because of code the first pass never showed. The lesson I’m keeping is about anchoring, because asking the agent for its lean before deciding quietly made the pick less independent.
Read: How do I know what I don’t know?
4. This Thursday: What your software factory is doing to your engineers
Software factories write a growing share of the code your team ships, and the review load lands on the few senior engineers with enough context to catch what’s wrong. In LinearB’s data from more than 8.1 million pull requests, AI-generated PRs merge at 33 percent against 85 percent for human-written ones.
On October 29 at 10am PT I’m hosting Marek Poliks, co-author of The Superdark Factory, and Cate Huston, author of O’Reilly’s The Engineering Leader, for a live workshop on where the software factory is headed. You’ll take away playbooks for measuring review load and team health in your own engineering data using LinearB.
5. AI;DR comes for the six-pager
A document used to be proof that its author had done the thinking, which was the idea behind Amazon’s six-pagers and the silent study hall at the start of meetings. Now a model can produce a convincing memo with no thinking behind it, and Steve Huynh, who spent almost twenty years at Amazon, describes a senior engineer in his community who tags AI-written docs AI;DR on the logic that if a model wrote it, a model can summarize it. He doesn’t think a ban is the answer. Instead he’d borrow Bezos’s one-way and two-way door framework, letting AI handle reversible decisions while documents like PR-FAQs, planning docs and promotion documents stay human-written. His second rule fits on a sticky note, “You own every word.”
Read: On the Death of Amazon’s Writing Culture
A Life Engineered6. Don’t build a fat bot
OpenClaw got a second look from Geoffrey Huntley, and his musings on building bots are the best part. Once a bot’s installed, the question becomes now what, and his answer is to stay curious, ask for what you want and turn whatever’s missing into a skill. He builds them thin, each doing one job and calling the others like Unix pipes, from an inbox bot that’s helped him unsubscribe from upwards of 300 newsletters to one that keeps tabs on unpaid conference reimbursements. We had Vincent Koc, chief architect at the OpenClaw Foundation, on the show last week, so if you haven’t listened yet, go check it out.
Read: revisiting some of my priors about openclaw
Subscribe today and add the one AI skill to your stack that never expires.
Subscribed© 2026 Dev Interrupted
548 Market Street PMB 72296, San Francisco, CA 94104