Connecting an assistant to this API takes about a minute. What you get is genuinely useful and genuinely limited, and the limits are not the ones people expect.
What it is good at
- One-off questions over data you would otherwise have to export. Who engaged with this post, and what do they do?
- Drafting, with you in the loop. It reads the thread and writes the reply; you decide whether it goes.
- Exploration. Finding out whether the data you need is even in there, without reading a reference first.
What it is bad at
- Anything that must happen on a schedule, reliably, without a person. That is a cron job, and a cron job is better at it.
- Anything where doing it twice is expensive. An assistant does not keep an idempotency table for you.
- Large batches. Hundreds of calls through a conversation is slow, expensive and harder to resume than a loop.
The boundary worth drawing
Use the assistant for questions and drafts, and write code for anything that repeats. That is not a limitation of this protocol; it is the difference between a conversation and a program, and it will still be true when the models are better.
The address you paste into the assistant carries your API key. An assistant calling with it reaches exactly what the key reaches — your connected accounts, and nothing else.
Questions people ask
Is a project room better than a single assistant?
Different. A single assistant is fastest for one person. A room adds what one assistant lacks: specialists nobody had to configure, other people reading along, and spending limits set once for everyone.
Can an AI agent in a room do more than an assistant outside it?
With the same tools, no. What it gains in a room is supervision, not power.