How HPE Zerto builds an agent system with Amazon Bedrock

The challenges of modern IT complexity

In the world of modern technology, IT infrastructures are becoming increasingly complex, and technical support teams face enormous challenges every day when it comes to troubleshooting critical systems. HPE Zerto , a leading platform in the field of IT resilience and data protection, decided to approach this problem in an innovative way, building an agent-based troubleshooting system based on Amazon Bedrock . This solution represents a significant step forward in automating troubleshooting processes and reducing resolution time for critical incidents. Agent-based systems represent a major evolution over simple traditional automation, as they can make autonomous decisions, execute chained actions, and interact with multiple data sources and external tools, simulating the thought process of a human expert experienced in troubleshooting complex data replication and recovery systems.

What is an agent system and why does it matter?

An agent-based system is an advanced artificial intelligence architecture in which one or more autonomous AI agents collaborate to solve complex, multi-step tasks without constant human intervention. Unlike traditional AI models that respond to a single query, AI agents can plan, use external tools, access knowledge bases, and iterate on results until they reach a defined goal. In the context of HPE Zerto , an agent-based troubleshooting system means that when a customer reports an issue, the AI ​​agent can analyze logs, query technical documentation, run automated diagnostics, and provide a complete, contextualized solution, dramatically reducing the reliance on human engineer intervention. This approach fundamentally transforms the way organizations manage IT incidents, turning processes that used to take hours or even days into a resolution experience that is measured in minutes.

Solution architecture: Amazon Bedrock at the heart of the system

Amazon Bedrock and foundational models

Amazon Bedrock is a fully managed AWS service that provides access to a wide range of foundational models (FM) from top suppliers such as Anthropic, Meta, Mistral AI and others, through a unified API. HPE Zerto chose Amazon Bedrock as the core of its agent system because of its flexibility, scalability and advanced capabilities. One of the major advantages of Amazon Bedrock is that it does not require infrastructure management for AI models, allowing engineering teams to focus on application logic and the quality of the end-user experience. In addition, Amazon Bedrock Agents provides a native framework for building AI agents that can orchestrate calls to external tools, manage conversational memory, and plan complex sequences of actions based on received instructions and available context.

The main components of the architecture

The architecture built by HPE Zerto includes several essential components that work together to provide an intelligent and efficient troubleshooting experience:

  • Amazon Bedrock Agents – the decision-making core of the system, responsible for orchestrating the workflow and calling the tools necessary to solve problems
  • Knowledge Base with RAG (Retrieval-Augmented Generation) – a knowledge base built from HPE Zerto technical documentation, support articles, troubleshooting guides, and history of previously resolved cases
  • Action Groups – sets of actions that the agent can execute, including querying monitoring systems, analyzing logs, and accessing internal APIs
  • Amazon OpenSearch Serverless – used for indexing and semantic search in technical knowledge bases
  • AWS Lambda – serverless functions that implement business logic for each action available to the agent
  • Amazon S3 – storage for technical documentation and logs used in the training and continuous improvement process

The agentic troubleshooting process: How it works in practice

Agent workflow

When a support engineer or customer initiates a troubleshooting session, the Amazon Bedrock agent takes control and begins a systematic investigation process. The first step is to gain a deep understanding of the reported issue by analyzing the initial message and asking clarifying questions if necessary. The agent then uses the ReAct (Reasoning and Acting) technique , an advanced paradigm in which the model alternates between explicit reasoning and the execution of concrete actions, allowing it to navigate complex troubleshooting scenarios with structured and transparent logic. This approach is superior to a simple chatbot because the agent does not just generate text, but actually makes decisions, executes operations, and validates the results before presenting a final solution to the user.

Integration with technical knowledge base

A central element of the HPE Zerto system is the RAG-based Knowledge Base . This technology allows the agent to retrieve relevant information from HPE Zerto’s extensive technical documentation, which includes hundreds of pages of configuration guides, troubleshooting articles, and release notes. The RAG process works in two main stages: first, the user’s query is transformed into a semantic representation vector, and then the system searches the vector database for the most relevant documents for that query. The retrieved documents are then provided to the foundational model as additional context , allowing it to generate accurate answers that are based on real technical information, rather than general knowledge or potential hallucinations. This approach significantly reduces the risk of incorrect information and increases user confidence in the solutions provided by the system.

Action Groups and integration with external systems

Action Groups are one of the most powerful aspects of the agent architecture. HPE Zerto has defined a comprehensive set of actions that the agent can execute autonomously to collect information and diagnose problems. These actions include:

System log analysis – the agent can request and analyze specific logs from the client's infrastructure to identify errors, warnings or anomalies that indicate the root cause of the problem

Component health check – the agent can query the health of various components of the Zerto solution, including Virtual Replication Appliances (VRA), Zerto Virtual Manager (ZVM), and Replication Journals

Configuration validation – the system can automatically verify that the current configuration complies with HPE Zerto recommended best practices

Identifying version conflicts – the agent can detect incompatibilities between the different components of the virtualized ecosystem

Generating diagnostic reports – at the end of the investigation, the agent can generate a complete report that documents the problem, the investigation steps and the solution applied

Technical challenges and how they were overcome

Managing context in long conversations

One of the major challenges in building an agent-based troubleshooting system is managing conversational memory during long, complex troubleshooting sessions. High-level technical issues can require dozens of message exchanges, the execution of numerous actions, and the analysis of large amounts of data. HPE Zerto addressed this challenge by implementing a sophisticated session management mechanism in Amazon Bedrock that allows the agent to maintain relevant context throughout the session, while eliminating redundant or irrelevant information to optimize the use of the foundational model’s context window. This optimization is critical because foundational models have a limited context window, and the efficient management of this resource directly determines the quality and consistency of the responses generated over an extended troubleshooting session.

Accuracy and confidence in AI responses

In the field of technical support for critical infrastructures, the accuracy of information is absolutely essential. An incorrect solution to a data replication problem can lead to data loss or unplanned downtime with severe financial consequences. HPE Zerto has implemented several mechanisms to ensure the quality of responses:

Grounding on verified technical documentation – all answers are anchored in official documentation, substantially reducing the risk of hallucinations

Citing sources – the agent always indicates the source documents that formed the basis of his recommendations, allowing engineers to validate the information

Automatic escalation – when the agent detects that the problem exceeds its capabilities or that the level of confidence in the solution is below a defined threshold, the system automatically escalates to a human engineer

Feedback loop – solutions validated by engineers are periodically integrated into the knowledge base, continuously improving system performance

Results and impact of implementation

Reducing resolution time

The implementation of the Amazon Bedrock- based agent-based troubleshooting system has generated significant results for HPE Zerto and its customers. The most notable impact has been a dramatic reduction in mean time to resolution (MTTR) for Level 1 and Level 2 incidents. Issues that previously required the involvement of an experienced engineer and could take hours or even days to resolve are now addressed in minutes through the intervention of the AI ​​agent. This improvement is possible because the agent can simultaneously process and correlate information from multiple sources, identify patterns in complex data, and instantly apply knowledge accumulated from thousands of previous similar cases, something that cannot be achieved with the same speed by a human engineer, regardless of their level of expertise.

Scalability of technical support

Another major advantage of the agent system is its unlimited scalability . Unlike human support teams, which are limited by the number of available engineers, the Amazon Bedrock-based system can handle an unlimited number of troubleshooting sessions simultaneously without degrading quality or increasing wait times. This capability is particularly valuable in crisis situations, when multiple incidents can occur simultaneously in HPE Zerto customer infrastructures. In addition, the serverless scalability of the AWS services used means that operational costs are directly proportional to actual usage, eliminating the need to provision excess capacity to handle peaks in demand.

Lessons learned and good practices

Designing agent prompts

The HPE Zerto team learned that the quality of the system prompts for the Amazon Bedrock agent is a determining factor in the performance of the entire system. Well-structured prompts should clearly define the agent’s role, its capabilities, its limitations, and how it should interact with users and available tools. A crucial aspect is to clearly define the chain of thought that the agent should follow when approaching a new problem, ensuring that the investigation is systematic and exhaustive before formulating a recommendation. HPE Zerto developed a detailed prompt engineering framework specific to the technical domain of data protection and replication, which guides the agent to ask the right questions, collect the necessary information, and structure the diagnosis in a way that is logical and understandable to end users.

The importance of data quality in Knowledge Base

The quality and relevance of the documents in the Knowledge Base have a direct proportional impact on the quality of the answers generated by the system. HPE Zerto invested significant resources in cleaning, structuring, and enriching the technical documentation before indexing it in the RAG system. The documents were organized hierarchically, relevant metadata was added to improve semantic search, and obsolete or contradictory information was removed. The Knowledge Base update process was automated to ensure that new product versions and new troubleshooting procedures are promptly reflected in the agent answers, thus maintaining the accuracy and relevance of the system over the long term.

The future of agent systems in enterprise technical support

HPE Zerto’s experience with Amazon Bedrock provides a clear vision of where enterprise support is headed. Agent systems represent not just an incremental improvement over traditional chatbots, but a fundamental transformation of how organizations manage technical knowledge and solve complex problems. In the future, we can anticipate the emergence of even more sophisticated agents, capable of proactive actions – identifying and resolving potential issues before they affect end users, by continuously analyzing telemetry from monitored systems. Also, integration with multi-agent orchestration systems , where specialized agents collaborate to solve problems that are beyond the competence of a single agent, will open up new possibilities for automating operational complexity in modern IT infrastructures. HPE Zerto has demonstrated that investing in agent technologies based on advanced foundational models is not a luxury, but a strategic necessity for companies that want to remain competitive and provide a superior experience to their customers in the AI ​​era.

HPE Zerto’s project to build an agent-based troubleshooting system with Amazon Bedrock is an excellent example of a practical and mature application of generative artificial intelligence in a real-world enterprise context, with high reliability and accuracy requirements. By intelligently combining Amazon Bedrock Agents , RAG technology, Action Groups, and AWS serverless services, HPE Zerto was able to create a solution that brings tangible value to both internal support teams and end customers. This implementation demonstrates that the future of technical support is agent-based, autonomous, and intelligent, and companies that adopt these technologies today will benefit from a significant competitive advantage in the long term. The model adopted by HPE Zerto can serve as a blueprint for other organizations that want to explore the transformative potential of AI agent systems in their own fields of activity.

Disclaimer:
This material was developed with the help of artificial intelligence for informational and educational purposes. The content was subject to human verification and review before publication. The information presented is intended to support the learning process and is not a substitute for consulting specialized sources, a specialist in the field, or participation in formal training courses and programs.