Short answer: Anis Bennaceur says deterministic workflows often work as well as or better than agents without heavy inference cost. Jordan Metzner says he has not found a use for agent SDKs beyond building agent-style tools.
Watch this answer from Episode 45 (starts at 22:40) · Episode 45 notes and transcript
From Episode 45: Your Pipeline Is Lying to You. This AI Fixes It After Every Sales Call, starting at 22:40.
Anis Bennaceur [22:39]: Yeah. I think there's two different things. The first one is how much your own product is consuming with these LLMs, and then how much the second thing is how much your team is consuming for all other operations that are set that are outside of your product if your product is an AI uses AI. Right? In the first case, it is extremely hard to kind of, like, track things properly. We have an expound tier guy here at our team, John Liu. He actually spent a lot of time just building something that understand how much each client is spending or costing us from an LLM perspective, how much each feature is costing us from an LLM perspective, and then how much each feature per client is costing us from an LLM perspective. So we've been able to track this really well now. And so that allowed us also to do a lot of optimizations on, like, which model to use and which cases and, from a latency and intelligence perspective. The second thing is from a, non product perspective.
Anis Bennaceur [23:40]: So from our operations perspective, this is where the caps are helpful. Right? Because Claude we're we're an Anthropic shop here. Claude will actually tell you within, its UI, this is how much you have for the day. This is left for the day. This is how much you have left until, your usage resets. And then if someone wants to use more, then they'll make a request. And either me or my cofounder will actually review it. It's not even one of the managers. It comes through me or the my cofounder, and we subjectively approve that. One thing that you will should be thinking a lot about more is, everyone's jumping on the bandwagon of agents. The reality is in lot of cases, workflows, just deterministic workflows tend to work as well, if not better, than stochastic agents. And so what you wanna do there is just, like, let your workflows rip. It will not incur any inference cost because it's just step by step things. Yes.
Anis Bennaceur [24:43]: Maybe some LLMs will be called, but it there won't be the heavy thinking that agents will do. Right? And so if you can actually automate a lot of your business through workflows rather than agents even in 2026, then I think there's a lot of gains that you can have there.
Jordan Metzner [24:57]: I totally agree with you. Every time I go to build something, I start with, okay, I'm gonna try to use the agent SDK, like, through OpenAI or Claude. And then as I, like, reason through it, like, we always create some type of deterministic workflow and use, like, just more of a classic LLM API calls and just kind of walk it through step by step to get the results rather than hoping the agent will do it. I honestly haven't found one use case where like we found the agent SDK to be useful for us other than building another agent SDK type of tool. Right? So, yeah, we're totally aligned. But, it's only one update away from seeing some step function improvements here. But alright, Anis. Well, this was a really fun episode. Great to chat. Love to learn about Attention. Obviously, you have a great domain name, so tell people how to find you.
Anis Bennaceur [25:40]: Yeah. Attention.com, best domain name in the world. You can also add me on LinkedIn. Message me. Tell me that you've seen me on this podcast. And yeah, I'll I'll respond there.
Jordan Metzner [25:50]: Awesome. Okay. Great. Alright, Sam. Well, this was a great episode. Obviously, another exciting week in AI and awesome to see some of the things that Attention is doing and definitely made me think a little bit about our SDR pipeline a little bit, what we gotta do to make some improvements.
Sam Nadler [26:03]: Likewise. Thanks, everyone. Thanks, Anis.
