Copy-paste is still king.

After setting up my own agents, I still find myself opening the Grok AI chat, pasting a paragraph and copying the result back into my work. I figured if I find myself falling back to this flow, then there's something special about it. The act of pasting text into a chat is fully controlled by the user, so there's no fiddling with invoked skills, a router that determines if a hook is a task, question, or note-taking.

Real case from a business owner

A engraving business gets design instructions from various channels - email, whatsapp, sms. We could plug an agent into each of these channels and act on incoming messages to trigger some automatic actions. The business owner would simply log on his computer and see first-pass generated designs ready on his filesystem. But this breaks apart in practice. Continuous conversation channels (whatsapp,sms) beget additions and corrections, but the agent would be triggered every incoming message. The business owner could reply, which adds important context to the conversation, but integrations don't capture the messages from the agent side. Additionally, the same customer could call in, or visit the store in person - all of this would be lost in the context capture. In the end, the business benefitted from adding durable instructions to his existing flow - this meant a simple system prompt to steer chat sessions.

So what are agents for?

Agents specifically benefit from strictly repeatable instructions. Anything that has to touch multiple domains - search, filesystem, cloud, and channels - but doesn't need a human making decisions at every step. It's easier to find where the agent usecase breaks than finding where the agent improves anything at all. But that's what makes finding those rare agent setups worthwhile - they are unique for every situation.