Zach Lloyd

So I was going to say, it totally does not need to write code.

But what's cool about having it actually be a coding agent is that it gives it a superpower.

So that if in the course of doing whatever you want it to do, it decides the best way to do it is by writing code.

It's really, really good at doing that.

And so...

Even for just normal knowledge work tasks, I find that these agents sometimes will be like, okay, actually the best way to do this isn't necessarily to just use the raw LLM response.

It might be to write myself a small Python script that scrapes websites and analyzes the data.

It's a pretty magical power when a coding agent does that.

Yeah, I think what we're seeing is that these agents are going into every function of the business, including our customer service, where we have a pretty significant problem with fraud and abuse where people...

script the creation of accounts to try to get higher free AI limits.

And then sometimes these people have the audacity when we block them to reach out to our support and be like, I can't believe you blocked it.

Exactly.

And so it's crazy.

And so it creates a pretty big support burden for dealing with people who in the first place are violating our terms of service.

And so we have put agents throughout this entire process of agents can sort of analyze the database.

They can look at the patterns of usage.

They can look at signals around IP and all other things that might indicate fraud.

Then by the time someone on our support team actually looks at one of these

inbound support requests, the agent has put together a whole risk profile on the person and an engineer didn't build this.

The person who built this is the person who is handling the support requests.