Submitted! I fall into a couple of the categories. I built a chrome extension that uses a streaming AI agent with n8n's MCP for tools. I combined it with a useful interface to manage and create ClickUp tasks while viewing emails, browsing online, etc. Honestly, AI agents become twice as useful when you have the right "interface" for giving them context and invoking them in the right moment.

I love using Gemini for video analysis! If anyone's interested, I have a template on the n8n site for analyzing Loom videos with Gemini (you just need a Loom video URL and it takes care of the rest). https://n8n.io/workflows/4472-ai-powered-loom-video-qanda-with-gemini-25-and-slack-notifications/

If you want to use it on your agent in n8n right now, you just need to manually switch the model to "claude-sonnet-4-20250514". I'm not noticing a big difference in behavior on my agents, but to be fair I didn't have intelligence issues with the types of tasks I was handling.

@Nate Herk That's pretty nifty. Have you run into issues with your agent hallucinating an email address? I told one to send an email to a specific email address and it misspelled it (I checked the logs to be sure). I was very surprised since the instructions were pretty empty, so it didn't have much to trip it up.

@Nate Herk What I want to ultimately do is build agentic AI with a combination of chat and UI interface. I don't love relying on written messages to see visual feedback of what my bot is doing. My ideal world is a combination of the bot confirming, and a UI card showing the email that was sent, calendar event booked, etc. It'll make hallucinations easier to catch. I'll have to play around with n8n AI more to see if I can do that with n8n + an external interface, or if n8n abstracts the raw data away and makes it unretrievable.

@Aj Silvers I've solved this problem before. Typeform's field definitions are separate from their answers array so you have to match them up by ID. Since this is a deterministic thing that you can solve with code and requires no judgment, I suggest using a code step to format the response in a useful way instead of using AI which is not only more expensive but also less reliable. Here's an example with your output and a code step to return an object with the key as the field ID, and the value as the user's answer. In the code you can also switch to using the field title as the key to make it a bit easier to find the right field, but if you ever change those questions then it'll break your chain so I recommend using the ID. Also note that in Typeform's output, the value of the answers has a different key for different types of questions. Multiple choice returns a "label" whereas text returns "text". There's a few other idiosyncratic ones. Hope that helps!