Short answer: Karim Malhas says every agent handoff adds time, so Breeze uses two or three agents at most for control, compliance or security, such as OTP checks done in deterministic code rather than inside a jailbreakable agent.
Watch this answer from Episode 42 (starts at 9:44) · Episode 42 notes and transcript
From Episode 42: The Voice AI Platform Built for Enterprises That Actually Need to Scale, starting at 9:44.
Sam Nadler [9:43]: Now I have a question. So I can totally see how you're, like, putting, like, incredibly powerful tools into operators' hands to build their own solutions, like, similar to n8n. How do you think, one of in our, brief use case of building voice agents, latency is, like, a huge challenge. How does this either, you know it does how do you balance the quality versus latency challenge?
Karim Malhas [10:06]: That is an excellent question. We've actually discovered that we cannot add too many agents because every time you're making a decision to go from agent to agent, it takes time. So we've discovered that we only use these when we need control, compliance, or security. So this agent might still come in and ask you, like, imagine this. Imagine you answer the call and instantly an API sends to your database or your server, which sends you an OTP, and this agent says, hey, just to verify, please tell me your OTP. Okay? This guy doesn't know anything about the OTP. So you may tell him what is the OTP. He doesn't know anything because it's in memory. So then its only role is to collect the OTP. And when it collects the OTP, I can then take it and do an if statement, which is code, so it's deterministic. And I if it's not, I go back. Ah, you didn't match. So I don't so now we have the layer of security that enterprises require and guarantee. Because I don't trust the agent with giving it the OTP and making the comparison within the agent, cause there is always a way to jailbreak it.
Karim Malhas [11:06]: I studied cybersecurity and I love to, like, kind of try to break things, and that you can break. But here, you can't. So you need to balance it. Initially, when I started to build this, I thought we're gonna build everything. Now I'm you're probably gonna need two, three agents tops because an agent is super good and life, especially when you're dealing with, like, any sort of situation, it's not deterministic like this and that or this. The humans and when you're ordering, for example, I may start and make an order, then I'm going to the end, and like, I'm done, and like, oh, no. I forgot something. So if you're gonna go back and forth, it's better to have it in a single agent. But there is situations where, like, I need consent and I cannot proceed, or I must get the OTP where this comes in handy. So this is kind of where you go. And now in terms of latency, you also have the models. This is why I love that we are agnostic and provide we don't care about we have this is a hundred milliseconds. This one is ninety milliseconds, and this is super fast. So you get the latency that you want. And I know this is kind of pivoting away from this voice agent, but it's a super important thing.
