Agent sovereignty, or why I don't want to depend on a single vendor
A €100 subscription started burning through a five-hour limit in 20 minutes. Since then I build the workflow so the model can be swapped and the work survives.
The subscription I pay €100 for started burning through its five-hour limit in 20 minutes of my usual work. The €20 subscription became decorative.
How we got here. Anthropic did a genuinely good job on their coding models, and people started buying subscriptions. Why not: every social feed carried the colourful story that you can build your own SaaS from a single prompt. OpenClaw got popular around the same time. And then it turned out that already loss-making subscriptions were consuming far too many resources.
Then came the changes. Claude CLI changed how it works with the cache and turned off important optimizations. That wasn’t enough — peak-hour limits got cut. Using your subscription from third-party tools was disabled too.
What did they tell users? Sorry about that.
The choice that was left
I had built my infrastructure on Claude CLI, and after all the changes I faced a fork: pay 200+ euros a month, or look for other options.
You’ll say: there’s an API key with no limits. There is. Have you seen the prices?
What annoys me separately is that some of the changes were never announced. Users complained en masse that the agent handled context worse and faded on long tasks, yet there was no official “nerf”. Numbers like 83.3% → 68.3% that spread through the chats back then are not Anthropic’s data — they’re results from third-party retests. The whole discussion was built on community observations, not release notes.
Then Anthropic started testing the removal of Claude Code from the Pro subscription, initially on 2% of users. The explanation was that subscriptions were never designed for multi-hour agentic tasks. OpenAI responded immediately: Codex access stays for Free and Plus alike.
What I mean by agent sovereignty
The idea is simple: my agents should work against an abstraction. When the next vendor starts tightening the screws — and that trend will continue and spread — I want to swap the model out without losing a workflow I spent months building.
In practice that means:
- the process lives in files and skills, not in a specific CLI;
- the spec, the plan, and the project rules sit in a repository, not in a chat history;
- any CLI can be replaced in an evening, because it holds nothing unique.
Yes, you’re still tied to somebody’s CLI — you could write your own, but why. The difference is that the choice of models inside gets much wider, and the cost of moving drops to zero.
Where it landed by May
Anthropic did acknowledge the problem eventually, from a different angle. Starting 15 June, every paid plan gets a separate monthly credit for the Agent SDK and claude -p — programmatic use stops sharing one pot with manual work. Before that you paid for Max, ran a swarm of agents through the SDK overnight, and discovered by morning that interactive Claude Code was already out of budget.
That’s exactly what the community had been writing on GitHub: claude -p billed against the subscription while the SDK did not, even though they do the same thing. Now at least the rules match.
The conclusion I drew earlier, and I’m not changing it. My skill set works with both Claude Code and Codex CLI. Not because the second CLI is better at anything, but because deciding “which one is cheaper to work on today” should never cost me my process.