What I Learned When My AI Agents Got Dumber Overnight
15 giu 2026
My best team members got dumb overnight. I had no say in it.
I run The Change Republic mostly on my own.
When last week Claude's newest model Fable 5 launched, I started using it the same day. In three days I built three assistants doing real work:
→ Expenses: 1.5 hours admin work per week down to 5 minutes
→ Newsletter workflow: fully automated drafting, publishing, design except my thinking
→ A new interface for my proprietary AI coach, easy enough for any client to use immediately (if you want to try: https://www.thechangerepublic.com/ai-coach)
I built all of this while submitting proposals, meeting clients, and doing all the regular mom things.
The model was sharp. Better than anything I had built before.
On day four, a US government directive suspended access for non-American users. Anthropic pulled it globally.
My "team members" did not disappear. They are still running. But the brain behind them switched back to the previous model overnight.
Same team. A lot less smart.
It was an aha moment for me. Until now I assumed the direction was simple. Tools keep getting better, and you always get the best one available (well, in commercial terms). This week was the first time I saw a brake on that.
A stronger model exists. I know, because I used it for three days. Right now I cannot access it. My agents will get sharp again the day it comes back, with nothing to rebuild. So I am not worried about that part.
But I realized, for the first time, the best AI I can use and the best AI that exists are not the same thing.
What would you do if your best team member got noticeably dumber overnight?
Want to think through what AI means for your own leadership? My Private AI Advisory is here.
What the best available model means for your AI strategy
In Private AI Advisory work with senior teams, the assumption I hear most often is that AI only ever gets better, and that you will always have the best version when you need it. Last week reminded me that this is not guaranteed. Access can be political, regional and reversible.
Three things this changes for leaders.
→ Build systems, not dependencies on one model. My assistants kept running when the model behind them changed. The work I had structured did not vanish. The structure is the asset, not the specific model underneath it.
→ Treat model access as a planning variable, not a constant. The best model that exists and the best model you can legally use in your region may differ. That belongs in your risk thinking, not as a surprise.
→ Keep the judgment human, so a model change is a change in speed, not in direction. When the thinking lives with you and your team, a quieter model slows you down. It does not steer you off course.
This is the difference between using AI as a tool and depending on it as a crutch. The leaders who handle weeks like this calmly are the ones who built for it in advance.
If this resonates and you want to think it through for yourself, I have a free AI Coach Tünde Light, a small set of self-coaching prompts you can use to slow down and take back agency before you hand a decision to AI. Leave your email here and I will send it over: https://www.thechangerepublic.com/free-resources