In Pegasystems' world, giving end users license to prompt agents can lead to trouble, except for research and summarization.
The latest agentic AI features for Pega's Infinity software suite give the prompting power to developers building the agents for users. Developers can use prompts, documentation, descriptions of automation goals, source code and videos of RPA processes to build workflow apps in Blueprint, Pega's agentic AI designer.
While other vendors such as Salesforce, Microsoft and ServiceNow rely heavily on prompts for their agents, Pega draws a line at giving end users prompts for tasks where there's a clear idea of the desired output, such as research that can be checked.
That kind of AI isn't needed to make decisions or handle complex workflows, said Predrag Jakovljevic, an analyst at Technology Evaluation Centers.
Agents that are primarily driven by prompts can go off the rails, using their AI reasoning power to take certain actions that the author didn't intend or anticipate.
Predrag JakovljevicAnalyst, Technology Evaluation Centers
"Agents that are primarily driven by prompts can go off the rails, using their AI reasoning power to take certain actions that the author didn't intend or anticipate," Jakovljevic said.
"Instead, Pega anchors its agents in workflows, which prescribe exactly how that agent should act at runtime," he said. "This makes the agents more predictable and reliable, which is particularly important in the highly regulated industries that Pega clients operate in."
Pegasystems CEO Alan Trefler said the agent workflow model is more reliable and predictable than the idea of prompts.
"The idea that somehow – that just with prompts we're going to reprogram our businesses and have them reliably do the right things, in a secure fashion -- doesn't play for most of what businesses do," Trefler said.
MCP standard making headway
Meanwhile, like its process automation competitors Microsoft, ServiceNow and Salesforce, Pegasystems has integrated Model Context Protocol (MCP) and Agent2Agent (A2A). In theory, these standards enable GenAI agents from different vendors to interact with each other for cross-platform automations.
The company has also released Pega Agentic Process Fabric, a library to orchestrate agents, apps, systems and data together. MCP can be an enabler for that, Trefler said.
Alan Trefler
Both Trefler and Jakovljevic acknowledged that the agent interoperability standards just came out earlier this spring. It will take some time for users to determine their effectiveness.
"More time and proofs of concept will be needed, "Jakovljevic said.
The technology is still in its early days, with many companies still only evaluating GenAI, let alone adopting its decision-making cousin, agentic AI. Different companies define agent differently by how they separate reasoning capabilities from data processing -- and check the agent's own work -- or by what decisions they are empowered to make.
"Pretty much anything that has an LLM in it -- or as part of the way it operates -- is actually calling those agents whether they have independent agency or not," Trefler said. "So to us, an agent is, in effect, a smart thing that accomplishes some outcome in a way that typically uses a language model."
Pegasystems' agentic tools were released in conjunction with its PegaWorld user conference this week in Las Vegas.
Don Fluckinger is a senior news writer for Informa TechTarget. He covers customer experience, digital experience management and end-user computing. Got a tip? Email him.
Alan Trefler
