Agentic AI Is Not Failing. We’re Just Judging It Wrong

For one month, Anthropic's Claude Sonnet 3.7, operating under the name Claudius, ran a small virtual automated shop inside Anthropic’s office. It:

It also made plenty of mistakes. It sold products at a loss, gave away discounts too freely, hallucinated a Venmo account, and at one point believed it was a man in a blue blazer delivering items by hand.

Eventually Claudius stabilised. It resumed its tasks and kept the store running. But it never turned a profit.

Because Claudius was not built to be flawless. It was built to test something new. Not just an assistant that responds to prompts, but an agent that makes decisions over time in the real world.

That shift changes everything.

You give it a task, it gives you an answer. If it fails, it’s broken. But agentic AI operates in uncertainty.

It reasons, adapts, makes decisions with incomplete information, and sometimes gets things wrong.

Not because it’s inadequate. Because it’s doing something fundamentally harder.

Judging Claudius by yesterday’s standards is like judging the first car for being slower than a horse or the first computer for taking too long to add. Paradigm shifts always look inefficient at first. They only make sense once we stop comparing them to what came before.

Claudius highlights how complex creating artificial 'agency' is. It needs memory, tools, planning, and the ability to revise its own strategies in motion.

We don’t yet have the infrastructure, workflows, or governance models to support systems like this. But that doesn’t mean they’re the wrong direction. It means we’re still early.

This is not the end of the story. It is the start of a shift, from assistants to agents, from tasks to decisions, from software that responds to systems that act.

The awkwardness we are seeing now is not a sign we are off track. It is a sign we are entering something entirely new.