What worked
I wired an agent into the real stack: Calendar, Drive, Gmail, and the local filesystem. It pulled a Stripe receipt for SuperGrok Plus and dropped the PDF in the invoices folder on my Mac. Not a slide. My desktop.
That is the bar I use now. If it cannot touch the tools I already pay for, it is a chatbot with a marketing team.
Where it stopped
Grok Bot is the one I was running. Same idea as the other agent tools: give it accounts, give it a job, and watch where it gets stuck. It got stuck on a domain purchase from a datacenter IP. It did fine on the invoice.
That split is the whole job. Filing an existing document was bounded, reversible, and easy to verify. A purchase brought identity, payment, anti-fraud, and ownership into the loop. Same agent. Completely different risk.
The operator takeaway
If you want to understand the wave, stop quoting model cards. Run an agent against one boring workflow you already know, then pay attention to the handoffs. The failure boundary will teach you more than the demo.