What exactly happened
On the morning of day four, the Anthropic API stopped responding with a 403 error. No prior email, no dashboard warning, no explanatory message. Access was blocked.
After contacting support, an answer arrived a few hours later: an automated system had detected an unusual API call pattern — high request volume in a short time, characteristic of agentic workloads. It was a false positive, but it caused a complete outage.
This isn't an isolated incident. Agentic usage patterns — large request counts, long sequences, repeated calls — look suspicious to security systems. And this doesn't only happen at Anthropic.
Why the orchestrator survived
The key decision from day two — the orchestrator must be separate from the specific model — turned out to be correct. The orchestrator didn't work with Claude directly; it worked with an abstract model interface.
Switching to the fallback model took about twenty minutes. Config change, basic functionality test, resuming work. No rewriting of orchestrator logic, no changes to the workflow.
If the orchestrator had been hard-wired to Claude, the experiment would have ended on day 4.
Single-provider dependency is technical debt
This experience convinced me: dependency on a single AI provider is an architectural risk in the same category as depending on a single database without backups or a single cloud region without failover.
A well-designed agentic system must be able to switch models at runtime:
- Provider outage — no cloud provider is 100% available
- Pricing changes — API prices change; the ability to switch saves your budget
- Quality updates — a new model may be better at a specific type of task
- Compliance — some organisations cannot send data to specific jurisdictions
The road to local models
The Anthropic ban accelerated a decision I'd been considering since day three: try local models as a primary or fallback option.
Local models don't have this problem. They run on your hardware or your cloud instance. No third-party provider, no automated security systems, no ban. Higher upfront investment, zero ongoing inference cost.
Day 5 was therefore dedicated to local models — and the results were surprising.
Want to see how business processes can be automated? Book a consultation — we start where vibe-coding ends.
In the next episode
Day 5: Three weeks with local models. Qwen, DeepSeek, OpenRouter and the return to Anthropic — with a pricing table.