The New Stack Agents – Antje Barth, Principal Developer Advocate, Generative AI, AWS

Key Concepts

• AI Agents
• Agentic Systems
• Foundation Models
• Strands Agents SDK
• Amazon Bedrock
• Amazon Q Developer CLI
• Model Context Protocol (MCP)
• Agent-to-Agent Communication (A2A)
• Guardrails
• Hallucinations
• Scalability
• Security
• Open Source
• Natural Language Interface
• Observability
• Traceability

AWS and AI Agents

• AWS's Approach: AWS aims to meet developers where they are, offering a broad and deep set of capabilities for building AI applications, from experimentation to production.
• Experimentation and Prototyping: AWS provides tools like Amazon Q Developer, an AI assistant in the IDE and CLI, to facilitate rapid prototyping and experimentation.
• Scaling and Production: AWS helps startups and developers scale their prototypes and onboard customers, offering support throughout the entire software development lifecycle.
• Competition: AWS focuses on providing a comprehensive stack of tools and services, emphasizing reliability and security.

Amazon Q Developer CLI

• Agentic Experience in CLI: Q Developer CLI brings agentic experiences into the terminal, allowing developers to interact with tooling using natural language.
• Natural Language Interaction: Developers can use natural language to execute CLI commands, access AWS environments, and learn new tools.
• Enhanced Developer Flow: Q Developer CLI helps developers stay in their flow by providing AI assistance directly within their development environment.
• Troubleshooting: It assists in troubleshooting errors by providing tips, tricks, and potential fixes.

The Rise of CLI-Based AI Tools

• Contextual Integration: CLI-based AI tools like Amazon Q Developer CLI, Gemini CLI, and Cloud Code bring AI assistance directly to the developer's workspace.
• Focus and Efficiency: By integrating AI into the CLI, developers can stay focused on their tasks without switching contexts or searching for documentation.
• Personalized Assistance: These tools act as personal tutors, providing information and guidance within the developer's existing workflow.

Bedrock and Agent Building

• Amazon Bedrock: A fully managed service offering a variety of foundation models through a single API, along with tooling for building agents and evaluations.
• Guardrails in Bedrock: Bedrock provides guardrails to ensure secure and controlled AI application development, mitigating factual errors and hallucinations.
• Automated Reasoning Checks: Bedrock uses automated reasoning checks to detect and correct factual errors stemming from hallucinations before responses are given to end-users.

Strands Agents SDK

• Open Source Initiative: Strands Agents SDK is an open-source project originating from internal AWS product teams.
• Simplified Agent Development: Strands simplifies agent development by leveraging powerful foundation models and reasoning models, reducing the need for extensive code.
• Model Context Protocol (MCP): Strands supports MCP, allowing agents to access a wide range of tools and capabilities.
• Name Origin: The name "Strands" was chosen by AI, representing the two strands of DNA connecting the model and the tools in an agentic system.
• Flexibility and Cost Control: Strands can be deployed locally, allowing developers to start building agents without immediate scaling needs, keeping costs in check.
• Model Integration: Strands supports custom model providers and integrations with Meta's Llama API, Anthropic, OpenAI, and LightLLM.
• Strandsagents.com: The website to find more information about Strands Agents SDK.

Strands vs. Bedrock Agents

• Choice and Flexibility: AWS offers both Strands Agents SDK and Amazon Bedrock agents to provide developers with choices based on their needs and preferences.
• Open Source vs. Managed Solution: Strands is an open-source project for developers who want flexibility and customization, while Bedrock agents offer a managed solution with integrated features and easier deployment.
• Use Cases and Evolution: Developers may start with Strands for learning and experimentation and then transition to Bedrock for more defined projects requiring a managed solution.

API-First Approach vs. Natural Language

• Shift to Natural Language: The industry is moving from an API-first approach to natural language interfaces, with English becoming a key aspect of generative AI.
• Agentic Services: The challenge is to transition from web services to agentic services, where agents browse applications and services instead of humans.
• Amazon Nova Act: Amazon Nova Act is a research preview project that helps models navigate browser interfaces designed for humans.
• Interface Redesign: There is a need to redesign interfaces to optimize them for agentic interactions, considering elements like pop-ups and deterministic vs. probabilistic systems.

Challenges and Mistakes in Agent Building

• Scalability: Building systems that can scale to millions of end customers is a major challenge.
• Latency: Reducing latency in multi-agent systems is critical due to the compounding effect of multiple model and agent calls.
• Optimization Techniques: Techniques like parallelization and caching are essential for optimizing agentic systems at scale.
• Multi-Agent Communication: Establishing communication between agents, both within and across companies, is a complex challenge.

Future of AI and Agents

• Pace of Innovation: The rapid pace of innovation in AI and agent technology is exciting, with significant potential for future breakthroughs.
• Democratization of Building: AI is democratizing the building process, making it easier for more people to start developing applications.
• Collaboration and Knowledge Sharing: The industry needs to come together to share knowledge, best practices, and address pitfalls in AI development.
• Advancements: The future may involve advancements in computer use, multimodality, and scaling agents further.

Security and Trust

• Importance of Security: Security and trust are critical considerations in AI development, requiring a threat modeling and threat analysis approach.
• MCP Security: AWS is contributing to open protocols like MCP to incorporate security best practices, such as authentication and authorization.
• Day One Consideration: Security and safety should be considered from day one of development, not as an afterthought.
• Traceability and Observability: Building traceability and observability into agentic systems from the beginning is essential for identifying and addressing issues.

Recommendations for Developers

• Open Source Communities: Engage in discussions within open-source communities and steering committees for security expertise.
• AWS Support: Reach out to AWS for help in building securely and safely, leveraging resources like the Gen AI Innovation Center and partner network.
• Community Engagement: Participate in community discussions and contribute to the implementation of security best practices in open protocols.

Strands Future Development

• A2A Support: The Strands team is actively working on full A2A specification support.
• Asynchronous Agents: Support for building asynchronous agents is being added to handle longer-running tasks without blocking.
• Public Roadmap: The Strands roadmap is public, encouraging community involvement and contributions on GitHub.

Synthesis/Conclusion

The discussion highlights the rapid evolution of AI agents and the tools and frameworks available for building them. AWS is positioning itself as a leader by offering a comprehensive suite of services, from the open-source Strands Agents SDK to the managed Amazon Bedrock platform. Key themes include the importance of scalability, security, and the shift towards natural language interfaces. The conversation emphasizes the need for developers to engage with open-source communities, adopt security best practices, and continuously learn and adapt to the evolving landscape of AI agent development.