Field Notes
When I wrote Don't Marry Claude, the point was simple:
Do not build your whole AI life inside one model provider's house.
Claude might be best today. GPT might be better tomorrow. An open model might be the right answer for a specific workflow. The model race moves too fast to make one vendor your foundation.
That idea was early in this series. A lot of you were not here yet.
So let me bring it forward.
AI sovereignty is going to matter more as agents move from fun demos into real business workflows.
The Fable/Mythos situation is the cleanest example I have seen so far.
In June 2026, Anthropic had to disable access to Fable 5 and Mythos 5 after a U.S. government export-control directive. Their public statement said the order applied to foreign-national access, and because they could not enforce that cleanly in real time, the practical result was that both models were disabled for all customers.
Other Claude models stayed available.
That detail matters.
A new capability appeared, people started building around it, and then the exact capability disappeared from the operating surface.
For a moment, people using Claude had access to what looked like the best model in the world for certain jobs.
Then the rules changed.
If your company had started building workflows around that exact level of intelligence inside Claude, you would have felt the trap immediately.
Claude was not the problem.
The brittle stack was the problem.
One of my own workflows hit this directly. Five sub-agents were assigned to run a phase of work. They were pointed at the model that no longer existed in the surface. The plan was clean. The documentation looked fine. The actual workflow hard-failed.
That is the difference between reading a leaderboard and running a business process.
A leaderboard tells you who is winning today.
A workflow tells you whether you can survive tomorrow.
Playbook
The lesson is bigger than "have a backup model."
Do not single-source your agent.
An agent in a real business is a workflow with context, tools, permissions, memory, review gates, and receipts.
Sovereignty has to cover the whole stack.
Start with the model layer.
You should be able to use the best model available this week without rebuilding your entire operating system next week.
That requires model routing. Keep the job description separate from the model. Keep your prompts, evaluations, and fallback rules somewhere you control. Know which jobs truly need the frontier model and which ones can run on something cheaper, faster, or local.
That is the practical version of Specified Intelligence. The best intelligence is the intelligence that fits the job.
Then check the context layer.
If your context only lives inside Claude, ChatGPT, or any other closed chat history, you are renting your memory.
That may be fine for casual work.
It is fragile for a business.
Your company brain should live outside the model: in documents, transcripts, notes, databases, tickets, SOPs, source files, and structured memory your team can inspect and move. The model should read the context. The business should own it.
That is why the open-stack conversation matters.
In The AI Tool Landscape, I broke down the difference between closed personal-agent surfaces and open personal-agent systems. Closed tools can be easier to start with. Open systems give you more control over the model, tools, memory, browser, files, scripts, and review gates.
You do not need to become a hardcore open-source purist.
You do need to know where your control lives.
Then check the workflow layer.
Your workflow should be able to move from one model surface to another without losing the whole process.
If your business process only runs because you know the magic prompt in one Claude project, you built a habit.
Useful, yes.
Portable, no.
The workflow needs to live outside the tool too.
Trigger.
Inputs.
Decision points.
Tools.
Permissions.
Review gate.
Definition of done.
Receipt.
That list is boring on purpose. It is also the part that makes the agent real.
In The New Software Stack, I wrote that the stack is becoming brain, body, tools, context, memory, review, and receipts. That stack now has a sovereignty test.
If any one layer owns the whole chain, you have a single point of failure.
Then check the tool layer.
Can the agent actually use the software, or does the tool only work when a person clicks through the screen?
Old-world software was built for humans staring at dashboards.
New-world software needs doors agents can use: APIs, webhooks, exports, CLIs, MCP servers, logs, scoped permissions, and clean audit trails.
This was the point of The Agentic Tool Test. A tool can be beautiful and still be a dead end for agentic work.
Then check the review layer.
If the same system that produces the work is the only thing reviewing the work, you have a blind spot.
For code, maybe Claude builds and Codex reviews.
For email, maybe the agent drafts and the human sends.
For pricing, maybe the system prepares the recommendation and waits for approval before touching the PMS.
For publishing, maybe the post cannot ship until the links, audio, layout, metadata, and email receipt pass.
Production and verification should be separate wherever the cost of being wrong matters.
That is sovereignty too.
Operational sovereignty.
You know what happened, who approved it, what changed, and where the proof lives.
Orientation
The lazy takeaway from the Fable/Mythos situation is "do not use Claude."
I would throw that answer out.
Use Claude.
Use ChatGPT.
Use Codex.
Use Gemini.
Use Grok.
Use open models.
Use whatever is best for the work in front of you.
The whole point is that you should be able to take advantage of the best model without letting that model own your business process.
Because this will keep happening.
The best model will change.
The cost will change.
The rules will change.
The interface will change.
The company will change its roadmap.
A regulator may change what a vendor is allowed to offer.
The people who win in this next phase will be the people who built flexible AI stacks.
Their context lives outside the model.
Their workflows live outside the chat window.
Their tools can be touched by humans and agents.
Their review gates do not depend on the same system that produced the work.
Their receipts show what happened.
So when the best model moves from Claude to ChatGPT to an open model to something none of us has heard of yet, they can move with it.
That is the advantage.
AI sovereignty lets you move fast without being owned by the thing that made you faster.
That is also why I am teaching this directly in my AI Bootcamp.
The current cohort is building real workflows now. The goal is an agent-ready operating layer: your context, your workflow, your tools, your review gates, your receipts, and your first custom agent.
That is exactly what I am helping people build.
On Friday, I am going to show you how to learn to execute everything we have been talking about across these first two arcs.
That email is the breakdown: how the waitlist works, what the next cohort is for, and how the first two arcs turn into the execution path for Arc 3.
Arc 1 was the mindset shift.
Arc 2 was the agent-ready stack.
Friday is the bridge into the build phase: the details, the next step, and what is coming as we move into AI engineering in plain operator language.
Comment below with the one AI tool or workflow your business would feel immediately if it disappeared tomorrow.
— Brian