Don Schuerman of Pega: LLMs + Agentic AI + Workflow Orchestration Bring Reliability and Predictability

Earlier this week Pegasystem announced the availability of Pega Agent Experience, a set of new API capabilities in its workflow automation and orchestration solution for delivering more reliable AI agents. I had the opportunity to speak with Pega Chief Technology Officer Don Schuerman to learn more about how it translates workflows into agents and how it works with Pega Blueprint, the generative AI-powered platform to streamline application and workflow design and creation.

In the clip below, Schuerman not only describes the how Pega Agent Experience and Blueprint work together take workflows created in Blueprint and transform then into agents that can communicate with processes and humans to get things done, but he also explains why the combination of workflows and agents provide more reliabilty and predictability than relying heavily on prompt engineering results.

Brent Leary: One other thing that that I picked up on that you said, and I think is really interesting,  was that this capability is kind of this one-two punch. It really allows the user to focus more on the workflow and not on becoming some kind of prompt engineer. Can you explain the importance of that?

Don Schuerman:We kind of think that in a lot of cases, the best thing to do with an agent is to actually get an agent to follow a workflow, to get an agent to actually follow the laid-out best practices that a business has agreed to.

Prompt engineering is really subtle, and I would say it's more of an art than a science. You can get very different responses from a large language model depending on how you phrase your prompt, how you use it. What model you're using on the back end can change how it interprets it, which is great if you're trying to do creative things.

It's great if you're trying to do some writing or you want to generate some images for you. Or like we do with Blueprint, to recommend the starting point for workflow definition. Oftentimes a business is certified with a regulator or certified with a compliance team rather than relying on just free-form prompts, which agents won't necessarily follow with any degree of predictability.

If I can use a workflow to tell an agent exactly what steps to follow, I get predictability. But I also get the power of the agent to be able to navigate that workflow, explain that workflow in different channels, interact with different people across that workflow.

I think there's a lot of enthusiasm around using agents. There's also a lot of uncertainty about how much of the decision making you actually want to leave to the agent, and how much you want to leave to the human being?

So those agent decisions and agent actions are predictable and [you can] explain them all vs. some of the natural unpredictability and randomization you've gotten from a lot of large language models. I think the big question becomes getting that to the point that it actually feels like something that fits within the predictable constraints of operating in a regulated enterprise.

To me this is a really exciting capability when you combine what's going on in large language models and agentic AI, and then you hook that up to workflow orchestration, we think you get something pretty cool.