In this session of the Predictable AI Placement series, we take a deep dive into the Step Agent—the third of the five AI placement patterns—and shows how it performs in‑process analysis to support faster, higher‑quality decision‑making.
Using a compliance and audit scenario, we demonstrate how the Step Agent is invoked during the audit assessment stage to perform a comprehensive analysis across multiple inputs. The agent evaluates auditor‑written findings, document summaries generated by the Doc Agent, audit objectives, and internal audit rules, and then enriches this analysis by calling an agent‑to‑agent (A2A) remediation service.
You’ll see how the Step Agent:
Runs automatically as part of the case flow (not conversational)
Consumes structured and unstructured audit data
Applies semantic analysis to assess findings against rules
Calls an external remediation agent to propose next steps
Produces a structured assessment summary written back to the case
The session walks through both runtime behavior and design‑time configuration, including:
Defining step‑agent instructions and guardrails
Using knowledge tools backed by data pages to retrieve case data
Integrating agent‑to‑agent calls for remediation suggestions
Writing agent output to a rich‑text field for human review and refinement
We emphasize how the Step Agent acts as a first‑pass analyst, accelerating assessments while keeping a human firmly in control. The results can be reviewed, edited, and approved before the case progresses, ensuring accuracy and accountability.
This deep dive builds on the earlier Application Agent and Doc Agent sessions and sets the stage for the next placement patterns—Assignment Assist Agent and GenAI Connect—completing the end‑to‑end picture of how AI placement patterns work together in Pega.
A user goes through a flow and has to wait for an Agent response without having initiated a conversation before
In the background, in parallell, an agent formulates a response
When the agent is ready, it starts a conversation with the user explaining the response and next steps
Is this an A2A-scenario too (Step-Agent followed by a Conversational Agent) and do you think it would be desirable to ‘pop-up’ a chat with the Conversational Agent instead of simply returning a response and letting users decide they want to have a conversation seperately?
That’s an interesting scenario , We can conditionally place a conversational agent in 3 places in a case, including a tab that will show the agent below the action area , in the utilities panel as well as the purple GenAI icon in the lower right. In terms of “popping up” dynamically when ready, we have not attempted that as of yet. At first glance , I see some possible patterns that could get close to that use case, but not exactly. We can conditionally show the agent in any of those 3 areas and also have a starter question that runs automatically. so we could have it as the default tab, with a starter question, ready to engage conditionally. Let me know if you are thinking of a specific business use case and we can build out a proof point for sure.