From autonomous AI Agents to predictable business outcomes: Pega’s approach to Agentic AI

AI agents promise a new step forward in the automation of business processes. Instead of merely generating information or supporting employees, agents can reason independently, take actions, and collaborate with other agents. This opens up interesting possibilities. But as soon as AI takes on a more active role in business-critical processes, an important question also arises: how do you give AI the freedom to reason without losing control over the process?

 

The risks of full autonomy

Fully autonomous agents may seem attractive at first glance. You give an agent a goal and then let it determine which steps are needed to achieve that goal. In a business environment, however, that freedom comes with risks. AI models can produce different outcomes based on the same or similar input. Agents may perform unnecessary actions, repeatedly call one another, or make decisions that do not comply with applicable business rules.

There is also a practical issue: every reasoning step and every interaction between agents consumes computing resources and tokens. When agents collaborate without clear boundaries, this can lead to unpredictable processes and unnecessary costs. The challenge, therefore, is not to give AI agents as much autonomy as possible, but to determine where that autonomy adds value and where control needs to be maintained.

 

Pega’s Predictable AI: autonomy with control

For organizations looking to use AI in areas such as customer service, financial services, or other business-critical processes, autonomy alone is therefore not enough. Processes must be predictable and secure, while also complying with applicable compliance requirements.

This is exactly the principle behind Pega’s Predictable AI approach. Rather than giving AI agents full control over a business process, Pega combines the reasoning capabilities of AI with the predictability of structured, deterministic workflows. This allows organizations to harness the power of AI within their processes while maintaining control over execution and working toward predictable, auditable business outcomes.

 

Letting AI reason within clear boundaries

The foundation of Pega’s approach is relatively simple: let AI do what AI is good at, while letting the workflow determine how the business process is executed. Instead of giving a generative AI model full control over an end-to-end process, Pega uses Predictable AI Agents for specific tasks within a predefined workflow.

Consider, for example, customer verification. The process itself can consist of fixed steps and business rules, while AI can be used within one of those steps to analyze information, compare data, or make a recommendation. Pega enables this with Agent Steps. These allow Predictable AI Agents to be embedded directly into specific stages of a structured case lifecycle. Within such a step, an agent is given a clearly defined task and only the context, data, and tools required to perform that specific task.

For this, Pega uses the open Model Context Protocol (MCP). Through MCP, the agent gains access to the information and tools it needs for its task, without automatically gaining control over the entire business process. That distinction is important: the AI reasons, while the workflow remains in control.

 

Human-in-the-loop

When a Predictable AI Agent does not have sufficient confidence in an outcome, an alternative path can be followed based on a predefined confidence threshold. For example, the workflow can then hand the task over to an employee for review and decision-making. Human-in-the-loop is therefore not a fallback solution added afterwards, but can be part of the process design from the outset. AI takes over repetitive or analytical work, while complex or uncertain decisions can remain under human control.

 

The Conversational Agent working together with the Application Control Agent

Predictable AI does not only change what happens behind the scenes. Agentic AI can also change how customers and employees interact with business applications. Traditionally, a user needs to know where to perform a certain task. Someone who wants to change their address, for example, first has to find the right application, navigate to the correct screen, and then fill out a form. With a conversational interface, this can work differently. A customer can simply say:

“I recently moved and would like to change my address.”

The user is not describing which steps the system should perform, but rather the outcome they want to achieve. Within Pega, the Conversational Agent plays an important role in this. It interprets the user’s natural language and understands what they are trying to achieve. As a result, the user does not need to know which workflow, application, or process steps are behind the request.

The Application Control Agent then comes into play. It has a different responsibility from the Conversational Agent. The Application Control Agent understands which workflows and capabilities are available within an application and uses an LLM in a targeted and controlled way to interpret the request, identify the relevant workflow, and map the required input.

In this example, the Application Control Agent recognizes that the request should be linked to the existing workflow for an address change and starts this workflow using the available data. Once the correct workflow has been started, the role of the Application Control Agent is largely complete. The workflow, rather than the language model, takes over the execution. Existing business rules, authorizations, security, process status, and controls therefore remain in effect. This makes it possible to create a more flexible user experience without giving the underlying business process the same degree of freedom.

 

Open standards as a connection between AI and workflows

An important part of this architecture is that Pega does not assume that agents only exist within Pega itself. Through the Model Context Protocol (MCP), compatible external agents and applications can also make use of capabilities made available within a Pega application. Consider, for example, agents that use Claude, Gemini, or OpenAI models. Such an external agent can describe a desired outcome, after which the Application Control Agent can identify and start the appropriate Pega workflow.

This creates an important distinction between interoperability and process control. MCP enables different AI systems to communicate with each other and with available tools and capabilities. Meanwhile, the underlying Pega workflow remains responsible for business rules, security, process status, and auditability. An external AI agent can therefore gain access to a business capability without automatically being given the freedom to determine the underlying business process itself.

 

Agentic Process Fabric: collaboration across applications

Many business processes do not stop at the boundary of a single application. Consider onboarding a new employee. HR needs to register the new employee, IT needs to set up accounts and provide equipment, identity management needs to assign access rights, and facilities may need to arrange access to the building. A manager could simply ask:

“Onboard Emma as the new marketing manager starting next Monday.”

Behind such a request are multiple workflows and applications. However, each Application Control Agent only knows the capabilities available within its own application. An additional orchestration layer is therefore needed. For this, Pega introduces Pega Agentic Process Fabric. This layer orchestrates agents and workflows across multiple applications. The Fabric determines which applications are needed, monitors dependencies and progress, and ultimately brings the results together.

In the onboarding example, each application used by the different departments can then execute its own controlled workflow, while Agentic Process Fabric coordinates the collaboration between those processes. To determine which workflows and capabilities are available, the Fabric uses an Intelligent Registry. Through Agent Cards, applications can describe the capabilities they offer, while the open Agent-to-Agent (A2A) protocol supports communication between agents.

 

From maximum autonomy to controlled autonomy

The discussion around Agentic AI often focuses on everything agents can do independently. For organizations, however, it is just as important to determine what an agent should not be allowed to decide independently. Which business rules must always be followed? What information is an agent allowed to use? When is human approval required? And can you reconstruct afterwards why a particular decision was made or action was taken? As AI becomes capable of performing more tasks, governance, security, and auditability become increasingly important.

Pega’s Predictable AI approach therefore does not aim for maximum autonomy, but for controlled autonomy. AI is used for tasks where interpretation and reasoning capabilities add value, while structured workflows remain responsible for the process as a whole.

With Predictable AI Agents and Agent Steps, AI can be deployed within specific process steps. With the Conversational Agent and Application Control Agent, users can initiate existing business processes using natural language. And with Pega Agentic Process Fabric, the same approach can be extended to processes that span multiple applications and agents.

 

Conclusion

This shifts the question from “How autonomous can we make our AI agents?” to a question that is ultimately more relevant for organizations: “How can we enable AI to do more while maintaining control over the outcome?” It is precisely this combination of intelligence and predictability that can enable the transition from experimenting with Agentic AI to deploying it responsibly in business-critical processes.