How distinct are all these new corporate "AI agents" under the hood?

I see AI agents everywhere now, from Instacart to Confluence to banking apps, and I was curious to know how they actually work behind the scenes. Are they mostly built on the exact same underlying models, or does each company code their own? If multiple companies use the same model, what stops them from acting exactly the same? Also, do they all benefit when the base model gets updated in the case they are running isolated versions? I'd love to know if they share any of the same training data or if each company's system is kept completely separate to what extent.