Note: I have 2 upcoming interviews as part of 60 Minute Stories that you won’t want to miss: Inessa Lurya discussing Building Zero-to-One Products in Established Companies and Sangeet Paul Choudary talking about Generative AI in Platform Ecosystems. Both events are free, live, virtual webinars…but seats are limited so sign up today!
One side benefit of these video interviews I’ve been doing is they give me a reason to call up former colleagues and old friends to talk shop. And in today’s episode, I share a recap of my conversation with my old Twitter buddy Arya Asemanfar. It’s not an exaggeration to say Arya was the tip of the engineering spear in the rearchitecture of Twitter’s platform services stack in the early 2010s; without his efforts, the Fail Whale would still be a thing. In short, he knows what he’s talking about, and when he’s hyped about something, it’s wise to listen. So my ears definitely perked up when he told me that agent development was the “next paradigm shift in software development”.
What are agents? And how does the build/test/deploy workflow differ from the traditional software development lifecycle (SDLC)?
Well, you can read more on that here and here in a couple of blog posts from Sierra (where Arya leads Product Engineering). But the reason this is a new frontier becomes clear when you hear him talk about how the traditional techniques and tools don’t translate over when dealing with a conversational paradigm vs a traditional web or mobile UI.
And with this new development scheme, there is also a new role emerging: agent engineer (which again the Sierra folks elaborate on here). This role is tasked with productionizing the power of LLMs in combination with orchestration logic such that businesses can replace workflows with agents. Arya and I talk through several customer support examples, since that is an obvious application of this capability (listeners and readers will recall my conversation with Anurag Wadehra from
where he talked AI being a boon to these types of workflows).In this next clip, Arya uses the transition from Photoshop to Figma as an analogy for how new agent design tools need to provide abstractions and primitives that allow agent engineers to be more expressive and design richer applications more quickly.
In fact, we’re so early in this journey that we have yet to publicly name these new patterns and interfaces. Arya makes a really interesting point about some of the most successful tech companies of the last 20 years not only building compelling products, but also creating the blueprint for how web, mobile, and services development happens; that’s exactly what’s happening with agent development right now.
We of course get into whether agents backed by LLMs are replacement or augmentation for today’s knowledge workers. The current consensus looks to be this will be a net improvement to the customer experience while also creating higher-leverage work for employees of business.
Another angle we delve into is the ability of agents to provide a more consistent experience as opposed to humans, who are more susceptible to variance in decisioning logic. It turns out predictability makes best practices (like experimentation) easier to introduce and business goals (like optimization) simpler to accomplish.
I’m including this next clip because in it Arya acknowledges that I asked a really “astute” question (it happens on occasion!). Basically, the rationale of an agent needs to be inspected from time to time in order to train and tune the system…but also so that humans can learn from it. Which in turn has led to the awareness that there is “an iceberg of necessary tools and products and experiences to effectively run an agent”.
“an iceberg of necessary tools and products and experiences to effectively run an agent”
We cover a wide variety of topics around agent development in the remainder of the conversation:
the need for escalation paths for when an agent’s business logic “hits a wall”
the likelihood that agents will fully replace humans in manual enterprise flows
agents interacting with other agents in a chained workflow that spans companies
