Your agent demos well. It answers cleanly in the notebook, handles the cases you thought of, and earns the budget to build the real thing. Then you try to ship it. Where does its memory live once the process exits? Who authorized the tool call it just made? None of that showed up in the demo. This book is about the second half of that story. It's the distance between an agent that runs on your laptop and one a business can depend on.

You start at setup and work toward the parts that decide whether an agent holds up under load. You'll structure agents you can see into, build with Strands Agents, LangGraph, and other SDKs, wire in real tools, and handle the errors those tools throw back. You'll coordinate more than one agent and keep the whole thing from getting brittle. By the time you reach MCP and Agent-to-Agent communication, the interesting question isn't what the protocols are. It's which one a given job needs, and how to run both together without the seams showing.

The rest runs on Amazon Bedrock AgentCore. A hotel concierge agent remembers a guest's preferences from one call to the next, because its memory outlives the session that created it. Tools and servers become callable endpoints through Gateway and get cataloged in Registry. AgentCore Identity authenticates every call, in both directions, with OAuth-based on-behalf-of tokens. OpenTelemetry traces land in CloudWatch, so when something breaks, the trace is there.

By the end you will have carried one system all the way across. It's deployed on AgentCore, monitored, and standing on the memory, identity, and gateway layers built earlier in the book. Whether you're moving an existing agent into production, starting from nothing, or building infrastructure a team will run, the work that keeps these systems alive is here.

You Will:

• Build and coordinate multi-agent systems with Strands Agents, LangGraph, MCP, and A2A, and know which protocol a given job actually needs
• Stand up production infrastructure on Amazon Bedrock AgentCore, with persistent memory, two-way OAuth identity, Gateway-exposed tools, and traces you can read in CloudWatch
• Keep the system alive after launch: watch its behavior, respond when it breaks, control what it costs, and stay compliant as load grows.

This Book Is For:

For engineers and architects building agentic systems. You should be comfortable with Python and have basic familiarity with LLM APIs. Some of job families include full-stack engineers, Data/Applied Scientists, AI Engineers, Enterprise Architects.

Ramcharan Kakarla is an Applied AI Scientist at Amazon Web Services (AWS), where he builds memory architectures and orchestration frameworks for multi-agent AI systems powered by large language models. He is the author of POC to Production: AI Agents in Action, and co-author of Applied Data Science Using PySpark (Apress), first published in 2020 and now in its second edition (2024). His work centers on agentic AI design and taking machine learning systems from prototype to production.

Before AWS, Ramcharan was a Principal Data Scientist and AI Engineer at Comcast, JPMorgan Chase, and Altice USA, shipping machine learning and agentic systems across telecommunications, financial services, and media.

He has published papers and posters on machine learning and AI, and served as a SAS Global Ambassador in 2015. He holds a master’s in management from the UCLA, and a master's from OSU with a specialization in data mining, along with a bachelor’s in electrical and electronics engineering from SASTRA University, India.

Raghu Rayasam brings over 18 years of expertise in architecting sophisticated artificial intelligence and data solutions. Currently serving as a Senior Data Engineer at AWS, Raghu specializes in building enterprise-scale AI systems, with particular emphasis on implementing advanced agentic workflows using AWS Strands, LangGraph, and AWS AgentCore. His innovative work in agent development encompasses robust memory management and agent observability through AWS AgentCore, enabling sophisticated multi-agent systems with enhanced monitoring and control capabilities. Raghu's expertise in integrating vector databases like OpenSearch with Bedrock's Large Language Models has set new standards in AI application development. Beyond his AI focus, he possesses deep expertise in distributed computing systems, including Apache Spark and Confluent Kafka, along with extensive experience in NoSQL technologies such as Cassandra and DynamoDB. His comprehensive background spans both on-premises and cloud environments, demonstrating his ability to bridge traditional data engineering with cutting-edge AI technologies to deliver transformative solutions that push the boundaries of what's possible in enterprise AI systems.

Dinesh Yadav Gaddam is an Applied Scientist at AWS whose work centers on generative AI and agentic systems. With a Master's degree in Management Information Systems from Oklahoma State University, he brings 12+ years of experience across industry leaders including Amazon, AWS, Comcast, and Cognizant, spanning Insurance, Media, Supply Chain, and E-commerce. Born and raised in Hyderabad, India, he completed his Bachelor's from Chaitanya Bharathi Institute of Technology (CBIT), Telangana. His research and applied work focus on fine-tuning large language models, applying reinforcement learning to align model behavior, and advancing prompt optimization techniques, with an emphasis on rigorous experimentation, evaluation, and measurable model performance. He designs, builds, and operationalizes LLM-powered agents that move from experimental prototypes to production-ready systems at scale. His contributions extend to applied research on temporal grounding for predictive recommendations, reflecting a commitment to translating scientific method into real-world impact across diverse business domains.