# Agentic AI
<a name="rise-agenticai"></a>

 **What is an agentic AI** 

Agentic AI refers to an autonomous AI system that can independently reason, plan, and execute complex, multi-step tasks to achieve a predetermined goal with minimal human supervision. Unlike generative AI, which primarily focuses on creating content based on human prompts, agentic AI is proactive and focused on taking action. It operates by continually perceiving its environment, reasoning through options, acting on its decisions, and learning from the outcomes in an iterative loop.

 **Types of agentic AI systems** 

Agentic AI can be deployed in different configurations, from single-purpose agent to large-scale multi-agent systems.
+ Single-agent: A single AI agent works alone to complete a defined, focused task.
+ Multi-agent: Multiple AI agents with specialized skills collaborate and coordinate to tackle complex workflows. This can be structured in a vertical hierarchy, with a lead agent overseeing others, or a horizontal, decentralized structure where all agents operate as equals.

 **Evolution into agentic AI** 

 **Stage 1:** More human oversight (Generative AI assistants) At the initial stage, AI systems primarily function as generative AI assistants, like early versions of chatbots or writing aids with high human involvement. It is reactive and prompt bases with “Human in the loop”.

 **Stage 2:** Generative AI agents This stage enhances the basic AI assistant with greater context awareness and tool-use capabilities, creating early generative AI agents with expanded capabilities with agents that are able to perform multi step tasks. They are governed by guardrails and still reliant on prompts.

 **Stage 3:** Agentic AI systems Agentic AI systems represent a major shift toward greater autonomy, integrating more complex reasoning, planning, and memory. They offer proactive execution instead on waiting on prompts, offer continuous learning, and with “Human on the loop” where the human role changes from direct involvement to strategic oversight.

 **Stage 4:** Autonomous AI agents The final stage involves the deployment of highly autonomous, multi-agent systems that operate with minimal human intervention. This has specialized multi agent collaboration to tackle complex end to end workflows and human focus shifts from oversight to governance.

 **Implementing agentic AI with Amazon Bedrock** 

 [Amazon Bedrock](https://aws.amazon.com/bedrock) provides a comprehensive and flexible toolset for building and deploying agents, supporting both fully managed and do-it-yourself (DIY) approaches. This is achieved by combining the fully managed and configuration-based [Amazon Bedrock Agent](https://aws.amazon.com/bedrock/agents/) with the highly customizable and composable services of [Amazon Bedrock AgentCore](https://aws.amazon.com/bedrock/agentcore/).

**Topics**
+ [

# Amazon Bedrock Agent
](rise-agenticai-bedrock-agent.md)
+ [

# Amazon Bedrock Agentcore
](rise-agenticai-bedrock-agentcore.md)
+ [

# Strands Agent
](rise-agenticai-strands-agent.md)
+ [

# Agentic AI to manage ERP Exceptions
](rise-agenticai-erpexceptions.md)

# Amazon Bedrock Agent
<a name="rise-agenticai-bedrock-agent"></a>

Amazon Bedrock Agent acts as the intelligent orchestrator that uses the reason-and-act (ReAct) pattern to fulfil complex user requests. It uses the reasoning of foundation models (FMs), APIs, and data to break down user requests, gathers relevant information, and efficiently completes tasks—freeing teams to focus on high-value work. You can refer to [this link](https://docs.aws.amazon.com/bedrock/latest/userguide/agents-how.html) on how to implement Amazon Bedrock Agent.
+  **User request**: The process begins with a natural language request from a user, such as "Generate a sales report and share it with the finance team".
+  **Reasoning and planning**: The Bedrock Agent’s orchestration prompt and the underlying FM interpret the request and break it down into logical, multi-step actions.
+  **Tool execution**: The agent executes the plan by invoking "tools"—action groups that are defined with API schemas. These tools can call backend services within the SAP system via the Generative AI Hub. For example, the agent might:
  +  **Call an API** to fetch sales data from SAP
  +  **Access a knowledge base** in Bedrock via a Retrieval Augmented Generation (RAG) tool to pull relevant business documents.
  +  **Leverage code interpreter or browser** in AgentCore for data analysis or to interact with a web-based SAP User Interface.
  +  **Utilize memory** to maintain context across multiple user interactions. This is essential for multi-step processes like filling out a complex purchase order over several turns of conversation.

Bedrock Agents fully supports multi-agent collaboration, allowing you to build and deploy systems of specialized AI agents that work together to accomplish complex, multi-step workflows. Instead of a single agent attempting to handle every part of a difficult task, a team of agents can be orchestrated to contribute their specific expertise, improving efficiency, accuracy, and overall performance. The core of [multi-agent collaboration in Bedrock](https://docs.aws.amazon.com/bedrock/latest/userguide/agents-multi-agent-collaboration.html) is a hierarchical model consisting of a supervisor agent and one or more collaborator agents.

# Amazon Bedrock Agentcore
<a name="rise-agenticai-bedrock-agentcore"></a>

Bedrock AgentCore is a suite of services that enables developers to build, deploy, and operate highly capable AI agents securely and at enterprise scale. It is designed to take on the "undifferentiated heavy lifting" of developing agentic AI, allowing enterprises to move beyond proofs-of-concept and accelerate production deployment. Bedrock AgentCore provides a modular toolkit of services that can be used together or independently to create sophisticated AI agents.
+  **Runtime**: A secure, serverless environment for deploying and scaling dynamic AI agents, supporting long-running and asynchronous tasks with complete session isolation.
+  **Gateway**: A service that converts existing APIs and AWS Lambda functions into agent-compatible tools with minimal code. It supports tool discovery and secure communication using protocols like Model Context Protocol (MCP).
+  **Memory**: Manages both short-term conversational context and long-term memory for agents, enabling more personalized and context-aware interactions without developers managing the underlying infrastructure.
+  **Built-in Tools:** Enhances agent capabilities with a Code Interpreter for secure code execution and a Browser Tool for interacting with web applications.
+  **Identity**: Provides a secure and scalable identity and access management service specifically for AI agents, integrating with existing identity providers to manage agent permissions.
+  **Observability**: Offers tools to trace, debug, and monitor agent performance in production, with comprehensive dashboards powered by Amazon CloudWatch and support for OpenTelemetry.

Bedrock AgentCore is explicitly designed to be model-agnostic, giving developers the flexibility to work with any foundation models (FMs) they choose, both inside and outside of the Amazon Bedrock ecosystem. These are some the FMs hosted within Bedrock, for full list you can refer to [this documentation](https://docs.aws.amazon.com/bedrock/latest/userguide/models-supported.html) :
+  **Anthropic**: The Claude family of models, including the latest Claude models.
+  **Meta**: The Llama family of models.
+  **Mistral AI**: A range of Mistral models.
+  **Amazon**: Amazon’s own models, including the Titan and Nova families.
+  **OpenAI**: Selected open-weight models from OpenAI.
+  **Other providers**: AI21 Labs, Cohere, DeepSeek, Stability AI, and others.

# Strands Agent
<a name="rise-agenticai-strands-agent"></a>

 [Strands Agent](https://strandsagents.com/latest/) is an open-source SDK created by AWS for building AI agents that use large language models (LLMs) to reason and act. The [Strands Agents SDK](https://github.com/strands-agents/sdk-python) simplifies the process of creating AI agents by focusing on three core components:
+  **A language model**: Strands supports a wide range of LLMs from providers like Anthropic, OpenAI, and Meta, giving developers flexibility.
+  **A system prompt**: This defines the agent’s role and overall behaviour.
+  **A set of tools**: These are the specific functions and capabilities the agent can invoke to perform tasks.

Benefit of strands SDK:
+ Strands SDK enables fast, secure development of advanced AI agents on SAP Generative AI Hub.
+ Developers can build complex automations quickly - saving time and resources.
+ Strands SDK supports multiple AI models and future technology shifts.
+ It has enterprise-grade security and robust monitoring ensure safe, reliable use.

![\[Strands Agent with Generative AI Hub and Amazon Bedrock\]](http://docs.aws.amazon.com/sap/latest/general/images/rise-agenticai-strandsagent.png)


The above architecture describes the integration option between Strands Agents, SAP Generative AI Hub to access Amazon Bedrock FMs, and Bedrock Agent SDK which allows integration to [Model Context Protocol (MCP)](https://www.anthropic.com/news/model-context-protocol) servers to access available APIs to automate workflows.

![\[Agent-to-Agent\]](http://docs.aws.amazon.com/sap/latest/general/images/rise-agenticai-a2a.png)


The most effective way in SAP is to have a Strands-built agent act as an external tool that an SAP Joule agent can call. This allows for specialized, custom logic to be developed in Strands, which is then orchestrated by SAP Joule within the business context of SAP applications. The architecture above describes how the [Agent-to-Agent](https://github.com/a2aproject/A2A) protocol works.

# Agentic AI to manage ERP Exceptions
<a name="rise-agenticai-erpexceptions"></a>

 **What is an ERP Exception** An Enterprise Resource Planning (ERP) exception is a notification generated by an ERP system when a real-world situation or process deviates from a planned norm, policy or rule. These exceptions act as alerts to indicate issues such as stock shortages, missed deadlines, or data discrepancies that require human intervention to resolve and prevent disruptions to business operations.

 **Why Agentic AI to manage ERP exception** Agentic AI goes beyond simply flagging an issue; it can autonomously reason, take action to resolve the issue, and learn from the experience. This moves ERP exception handling from a reactive to a proactive and preventative process.

 **How agentic AI improves ERP exception handling** 

Agentic AI to manage ERP exception handling helps with

1. Proactive problem-solving

1. Faster and more autonomous resolution : Agentic AI can resolve many exceptions without human intervention by learning from historical resolutions

1. Continuous learning and improvement

1. Intelligent routing and escalation

1. Enhanced compliance and auditability since every action taken by an Agentic AI agent can be audited and guarded with guardian agent

1. Freeing up human resources

 **Top use cases for ERP exceptions management with Agentic AI** 

 **Use Case 1: Three-way Invoice Matching** In this process, we match Purchase Order against Goods Receipt and Invoice. The exception cases of unmatched invoices are sent to the AI agent. It does the same research that the user would have done, the AI agent successfully finds the correct PO number, saving the exceptions user the time of doing the research. The exceptions user reviews the Agent’s findings and approves. The agent processes the transactions saving the exceptions user the time of processing the transactions.

 **Use Case 2: Customer Payment Matching** In this process, we match the invoice against customer payment in bank statement. The exception cases (unmatched customer payments) are sent to the Agentic AI Agent. The AI Agent does the same research that the user would have done. It will find the invoice and match to the customer payment from bank statement and presents the recommended solution to the user, saving the user the time of doing the research. The exceptions user accepts the recommendation. The agent processes the transactions saving the exceptions user the time of processing the transactions.

 **Use Case 3: Sales Order Entry** In this process, a certain sales order line item has no available stock to fulfil. The Agentic AI Agent retrieves information from the ecommerce site, emailing the customer with a replacement SKU and escalating to the credit and supply chain team. After completing the research, the agent will recommend a solution for each exception. If the user accepts the recommendation, the agent performs the transactions in SAP and/or other systems to replace the item.

 **Use Case 4: PO Confirmation** The Agentic AI Agent can parse each PO to extract key terms such as limits of liability and compare the key terms with the central contract automating the PO Conformation Process. Upon confirmation, the Agent can enter the PO as an order into the ERP system.

 **Use Case 5: Cash Forecasting** The ERP system contains most or all information required for creating a cash forecast. The ERP has bank account balances, unpaid vendor invoices, unpaid customer invoices, and other critical inputs to a cash-forecasting process. Other systems may also contain additional information for input into the cash forecast. A forecast is generated from bank/investment account balances, vendor invoices (liability) and customer invoices (asset). The Agentic AI Agent collects the necessary data points from the ERP and other systems and calculates a per-day cash forecast based on standard operating procedure.

 **Use Case 6: Financial Period End Close** In this process, AI Agent can do several, most or even all of the steps for financial period-end closing with or without a human in the loop. The Agent can reconcile bank statements, account receivables and payables, consolidate ledgers, and account for depreciations, unearned revenue, prepaid expenses and intercompany reconciliations. It can handle the exceptions by communicating with various stakeholders in the organization.