Downloaded 10 times
Uploaded byAWS Chicago
Alec MacEachern - Scaling Enterprise Agents
AWS Community Day Midwest 2025 Alec MacEachern Scaling Enterprise Agents
More Related Content
Similar to Alec MacEachern - Scaling Enterprise Agents
![[DSC Europe 24] Lana Malic - AI Agents: The Future of Autonomous Decision-Mak...](https://cdn.slidesharecdn.com/ss_thumbnails/lanamalic-aiagents-250212223710-84219c4c-thumbnail.jpg?width=640&height=640&fit=bounds)
Alec MacEachern - Scaling Enterprise Agents
- 1.
- 2. Scaling Agents with EnterpriseData Alec MacEachern Head of AI, Uturn Data Solutions
- 3. • Headquartered inChicago, IL • Founded in 2013 • 100% AWS Focused • Advanced Tier Partner – Since 2014 • Well-Architected Program Launch Partner • Near-Perfect AWS Customer Satisfaction Rating • Servicing customers around the world We are obsessed with helping our customers be successful with public cloud. We LOVE data. Uturn: Who We Are • Migration Competency • DevOps Competency • Data and Analytics Competency • AWS Lambda Delivery Partner • Amazon Redshift Delivery Partner • Amazon EC2 for Windows Delivery Partner • AWS Marketplace Partner • Well-Architected Partner Program • Immersion Day Partner • MAP/MAP Lite Funding Approved • Migrate IW Funding Approved • Data & Analytics IW Funding Approved • Data Foundations for GenAI DATA
- 4. o Introduction toAI Agents and their capabilities o Recent advancements in agentic frameworks and tooling through OSS and AWS o Architectural patterns to discuss around orchestration, scalability, security o Enterprise use cases o Discussion/Action Items/Q&A Agenda
- 5. • Autonomous entity:Software agent that perceives its environment, reasons, and acts to achieve goals. • LLM-powered tools: Breaks user requests into steps, uses knowledge bases or APIs (tools), and composes responses. • Agents can* act like a smart assistant: Incorporates memory, planning, and decision-making to automate complex workflows beyond simple chatbots. * When given agency for self-directed operation in its tasks that it is grounded to What is an AI Agent? Source: Langgraph docs: https://langchain-ai.github.io/langgraph/concepts/multi_agent/#multi-agent-architectures
- 6. Cognitive Load Reduction •Each agent focuses on narrow domain vs juggling multiple contexts Prompt Optimization • Domain-specific prompts vs generic "do everything" prompts Context Efficiency • Only relevant knowledge loaded vs broader context dilution Why Multi-Agent Systems vs Single Agent Chatbots? Tool Selection Accuracy • Pre-selected relevant tools vs choosing from larger tool set Decision Simplicity • Clear routing upfront vs constant internal mode switching Specialist vs Generalist • A cardiologist will outperform a general practitioner on heart issues because of focused expertise, even though the GP might know some cardiology
- 7. MCP (Model ContextProtocol) • Anthropic’s Open standard for connecting LLMs to 3rd party tooling and data ingestion resources • Defines roles (host, client, servers) so agents can call APIs and use data uniformly • Abstracts away model-to-tool integration by having tools self-describe inputs/outputs OK, but how do we start with Agents? A2A (Agent to Agent) Protocol • Google’s Open stanard for agent interoperability • Enables AI agents from any model provider to communicate capabilities, assign tasks, and coordinate actions across enterprise systems
- 8. • Removes thebottle neck of needing to interface LLMs with multiple data and tooling integration that are prevalent across the enterprise • Leverages Client-Server architecture for local- first connections for more secure data transactions and permissions granting • Every major cloud provider and OpenAI have MCP integrations, which will standardize MCP for their native service offerings MCP for Tooling Integration Source: Conor Kelly: https://humanloop.com/blog/mcp
- 9. • Creates clearcommunication channels between disparate agentic systems • Enables agents to delegate tasks, exchange context, and work together on complex user requests. • Facilitates secure communication patterns suitable for enterprise environments, relying on standard web security practices. • Reuse existing, well-understood standards (HTTP, JSON-RPC 2.0, Server-Sent Events). A2A for Agentic Distribution
- 10. • OWASP’s AgentName Service is addressing the lack of a public agent discovery network (DNS for agents) • Protocol-agnostic registry that use PKI certs to verify agent identity and trust • Provides vendor agnostic registries for enterprise scaling of agentic marketplace adoption How can we secure agentic proliferation?
- 11. • ANS actsas a core infrastructure function • Enables Communication post authentication • Integration of control layers for oversight and coordination during agent discovery processes ANS Architecture Example Source: Reuven Cohen’s repo: https://github.com/ruvnet/Agent-Name-Service/blob/main/docs/architecture/ ans_architecture.md
- 12. • There areno agents with agentic frameworks for orchestration. • Let’s look at some options, both OSS and offered through Bedrock How do we manage these systems?
- 13. • Integrations toany major cloud LLM provider, MCP servers for tooling and datastores, vector databases, etc • Langchain helps with connecting these disparate tools to work as a single system • LangGraph for agentic orchestration • LangSmith for GenAI Ops components Langchain Components
- 14. • Orchestration layerfor application integration, prompt + model management, and tools integration. • Really great for .NET developers and Microsoft ecosystems + integrations with Amazon Bedrock Semantic Kernel
- 15. • AWS Supportfor pro code and low code tooling to develop and deploy agents • Great for starting off while you build and design supervisory agentic systems Bedrock Agents
- 16. What is it? •The InlineAgent is a sample implementation demonstrating how to define and run an Amazon Bedrock Agent entirely in code (i.e., “inline”) without relying on external configuration files. What it does: • Define agent instructions and tools programmatically. • Use MCP to connect the agent to tools. • Run the agent locally using the AWS SDK (boto3). Inline Agent SDK In Bedrock
- 17. • AWS’s OSSframework for building, deploying, and managing agents • An agent in Strands engages in an iterative loop, reasoning, selecting tools, and executing actions with its model until the task is fully resolved. AWS Strands
- 18. • Prior Authis a very complex workload between provider and payor systems that takes a lot of manual work. • Both sides of the system can be automated to expedite the clinical data classification/extraction pieces and assist with the submission and context for decision making criteria Prior Authorization Use case
- 19. Investment Agents Examplein Bedrock Agents
- 20. • This isstill very early days of Agentic Systems! • 2 of the major components discussed were just released in the last couple of months, and hardly any of this existed a year ago. • Go and check out curated lists of MCP servers. They are published all over Github. • Feel free to add me: www.linkedin.com/in/alecmaceachern
- 21. Questions? Please ask. AlecMacEachern [email protected] Thank you uturnData uturn-data-solutions
Editor's Notes
- #11 Core Infrastructure Function: ANS serves as the foundational identity management system for the Mastra ecosystem, functioning like DNS by maintaining a registry that connects agent identifiers with their corresponding network endpoints and security credentials. Communication Enablement: The service facilitates secure direct communication between agents (A2A) by providing the necessary routing and authentication data required for establishing trusted connections across the network. Management Integration: ANS seamlessly integrates with the Management Control Panel (MCP), allowing for centralized oversight and coordination of agent discovery processes and communication pathways within the distributed system.
- #13 LangChain Unified interface to model selection and data stores to chain tools to agent for sophisticated enterprise needs Offers both pro-code and low code approaches
- #16 Inline Agent This example is great for developers who want to build and test agents without setting up full infrastructure. It’s a hands-on way to understand how Bedrock Agents work under the hood, especially with MCP and inline tool definitions.