How can I move agentic workflows from my laptop to production?

Productionizing agentic workflows involves using managed, serverless services like AWS Lambda for tools and Amazon Bedrock for the LLM and agent orchestration. These services handle scaling, security, and infrastructure management.

What is Amazon Bedrock and how does it relate to agents?

Amazon Bedrock is a fully managed service that provides access to various foundation models and offers a managed platform for configuring, deploying, and scaling agents.

How do agents connect to external systems or functions?

Agents connect to external systems through 'action groups,' which can contain functions like AWS Lambda. The agent is configured with a function schema that describes these tools in natural language, allowing the LLM to reason about and use them.

What is the purpose of the code interpreter in Amazon Bedrock agents?

The code interpreter allows agents to write and execute code (e.g., Python) in a secure, managed environment. This is crucial for performing deterministic tasks like mathematical calculations or data manipulation that LLMs are not inherently good at.

How do Guardrails enhance the safety of agentic workflows?

Guardrails monitor both the input prompts to the LLM and the output from the agent. They act as a safety mechanism to prevent prompt attacks, filter out harmful content (like hate speech or insults), and ensure the agent's responses align with defined policies.

What is Retrieval Augmented Generation (RAG) and how does Bedrock Knowledge Base help?

RAG allows LLMs to access external data beyond their training set. Bedrock Knowledge Base simplifies RAG by managing the process of ingesting documents, chunking, embedding them into vectors, and storing them in a vector store (like OpenSearch), making the data retrievable by the agent.

What are Amazon Nova models?

Amazon Nova models are a family of high-performance and cost-effective LLMs developed by Amazon, available through Bedrock, offering an alternative to third-party models.

What are inline agents in Amazon Bedrock?

Inline agents allow you to invoke agent configurations directly without deploying a fully managed agent in the cloud. This speeds up development and testing, enabling you to iterate on agent configurations rapidly.

What is 'return control' in the context of agents?

'Return control' refers to a pattern where the agent identifies the next tool to use but returns control to the developer or another system to execute that tool. This is useful for local execution or when managing agent state outside of the cloud.

Key Moments

AI Dev 25 | Mike Chambers: Serverless Agentic Workflows with Amazon Bedrock

DeepLearning.AI

Entertainment3 min read64 min video

Mar 27, 2025|1,664 views|30|3

Save to Pod

Key Moments

TL;DR

Build scalable, secure serverless AI agents with AWS Bedrock, integrating tools, code interpreters, and knowledge bases for production.

Key Insights

Amazon Bedrock provides a managed service for building and deploying AI agents, abstracting infrastructure complexity.

Agentic workflows leverage LLMs and tools (like Lambda functions) to interact with external systems and perform tasks.

AWS services such as Lambda, S3, DynamoDB, and OpenSearch are integral to building robust and scalable agentic systems.

Key features for production-ready agents include managed tools, a code interpreter for deterministic tasks, guardrails for safety, and knowledge bases for data access.

Newer features like inline agents and Amazon's Nova models offer faster development cycles and enhanced capabilities.

Serverless architecture is crucial for scalability and managing infrastructure without manual intervention.

INTRODUCTION TO AGENTIC WORKFLOWS AND PRODUCTION DEPLOYMENT

The session introduces agentic workflows, defined as systems that use generative AI to connect with external systems. The primary focus is on moving these agents from local development to production environments, emphasizing the importance of scalability, security, and manageable infrastructure. This transition is a growing concern for organizations looking to realize tangible outcomes from their AI investments.

AMAZON BEDROCK AND AWS SERVICES FOR AGENTS

Amazon Bedrock serves as a fully managed service for building generative AI applications, offering access to a variety of foundation models from different providers, including Amazon's own Nova models, Anthropic, and Meta. It allows seamless integration with other AWS services like Lambda for running code snippets at scale, S3 and DynamoDB for data storage, and managed components for agent configuration, retrieval augmented generation (RAG), and guardrails.

CORE COMPONENTS OF AN AGENTIC WORKFLOW

An agent acts as an orchestrator powered by a large language model, utilizing tools (represented as external systems) to interact with the outside world. Key to agentic workflows is the ability to self-loop, allowing the agent to chain tool executions and reasoning steps. For production, these agents can be configured and managed through Bedrock, ensuring scalability and reliability without manual infrastructure management.

ENABLING AGENTS WITH TOOLS AND FUNCTIONALITY

To extend agent capabilities, developers can define 'action groups' that map to external services, such as AWS Lambda functions. Each action group contains specific actions (tools) with detailed schemas describing their purpose and parameters. This allows the LLM, through the agent, to understand and invoke these tools, enabling it to perform complex queries, retrieve data, or trigger backend processes like customer support interactions.

ADVANCED CAPABILITIES: CODE INTERPRETER AND GUARDRAILS

For tasks requiring deterministic execution, such as mathematical calculations or data manipulation, Bedrock agents can be equipped with a code interpreter. This secure, managed environment allows the agent to write and execute code, returning results to the LLM. Additionally, guardrails are implemented to monitor input and output, ensuring policy compliance, preventing prompt attacks, and maintaining model safety and reliability, crucial for customer-facing applications.

KNOWLEDGE BASES AND RETRIEVAL AUGMENTED GENERATION (RAG)

Fully managed knowledge bases enable agents to access proprietary data. Documents are stored in S3, then processed by Bedrock Knowledge Base to be chunked, embedded into vectors, and stored in a vector store like Amazon OpenSearch Serverless. This integrates retrieval augmented generation (RAG) capabilities, allowing agents to provide informed responses based on up-to-date company information, greatly enhancing their utility for tasks like customer support.

INNOVATIONS: AMAZON NOVA MODELS AND INLINE AGENTS

Recent advancements include Amazon's own Nova models, offering improved price-performance for generative AI tasks. Furthermore, the introduction of 'inline agents' simplifies development by allowing agent configurations to be invoked directly within a development environment without full cloud deployment. This approach accelerates iteration cycles and opens possibilities for dynamic agent configuration generation and local tool execution.

PRODUCTION READINESS AND SERVERLESS ARCHITECTURE

The entire architecture discussed emphasizes serverless components, meaning there is no infrastructure to manage, patch, or scale manually. Services like Lambda, Bedrock Agents, and managed databases automatically scale with demand. This serverless approach is fundamental to achieving production-ready AI agents that can reliably handle varying workloads and deliver consistent performance.

Mentioned in This Episode

●Products

●Software & Apps

●Organizations

●Books

●Concepts

Key Steps for Deploying Serverless Agentic Workflows

Practical takeaways from this episode

Do This

Leverage AWS Lambda functions for agent tools, benefiting from scalability and security.

Utilize Amazon Bedrock for foundation models and its managed agent platform.

Configure agents with clear instructions, roles, and an access model ID.

Prepare agents by running the 'prepare' command and wait for the 'prepared' status.

Create an alias for your agent, similar to Lambda aliases, for production-ready access.

Use the Bedrock Agent Runtime client for invoking prepared agents.

Enable traces to understand agent reasoning and for debugging.

Define Action Groups to connect agents to external systems like Lambda functions.

Provide detailed descriptions in function schemas for the LLM to understand tool capabilities.

Enable the Code Interpreter tool for agents to write and execute code.

Integrate Guardrails to monitor agent inputs and outputs for safety and policy compliance.

Configure Bedrock Knowledge Base to connect agents to your data sources (vector stores like OpenSearch, S3).

Explore Amazon Nova models for cost-effective, high-performance LLM capabilities.

Use inline agents for faster development and testing cycles by invoking configurations directly.

Consider 'return control' for scenarios requiring local tool execution or persistent agent state.

Avoid This

Do not assume an agent can interact with external systems without defining tools and action groups.

Do not rely solely on the LLM for deterministic tasks like complex mathematics; use the code interpreter.

Do not forget to prepare and alias your agent before invoking it in a production-ready manner.

Avoid complex, ad-hoc setups for RAG; use managed services like Bedrock Knowledge Base.

Do not underestimate the importance of clear descriptions in function schemas for the LLM's understanding.

Be cautious with prompt attacks; utilize Guardrails to mitigate risks.

Do not attempt to deploy complex or untested changes without proper testing, especially in production.

Common Questions

An agent is a system that uses generative AI to connect with external systems and tools beyond its internal model capabilities. It acts as an orchestrator that can loop back to itself, interact with tools, and reason through a process to achieve a goal.