Building Production AI Agents with the Strands Agents SDK for TypeScript

Overview of Strands Agents

Model-Driven Approach to Building Agents

• Strands takes a model-driven approach to building agents, where the agent is given a goal and a set of tools to achieve that goal
• This is different from a workflow-driven approach, where the developer has to define the specific steps the agent should take
• With the model-driven approach, the agent is able to figure out its own workflow to achieve the goal, making the development process much simpler

Key Features of Strands Agents

• Supports a wide range of reasoning models, including large language models (LLMs) from providers like OpenAI, Anthropic, and Cohere
• Allows developers to bring their own custom model providers and tools
• Provides a simple, intuitive interface for building agents in just a few lines of code
• Includes features for controlling and steering the agent's behavior, such as structured output, system prompts, and hooks

Strands Agents SDK for TypeScript

Motivation for Typescript Support

• The Strands team wanted to provide TypeScript support to improve the developer experience, building on lessons learned from developing the Python SDK
• They aimed to make the TypeScript SDK high-quality and maintainable, rather than relying heavily on AI-generated code

Development Approach

• The team used a collaborative approach, working with the Agents team to develop the TypeScript SDK
• They leveraged Strands' own features, such as agent SOPs (Standard Operating Procedures), to guide the agent in writing the TypeScript code
• The team also integrated the TypeScript SDK development into their GitHub workflow, using Strands agents to automate code generation and updates

Key Features of the TypeScript SDK

• Supports the same simple agent loop as the Python SDK, allowing developers to quickly build prototype agents
• Includes support for Bedrock and OpenAI models, as well as the ability to bring custom model providers
• Provides features for controlling and steering the agent's behavior, such as structured output and system prompts
• Integrates with the Strands SOPs (Standard Operating Procedures) system for more structured agent guidance

Improving the Developer Experience with Strands

Agent SOPs (Standard Operating Procedures)

• SOPs are a way to provide structured, step-by-step instructions to the agent, using keywords like "must" and "should" to guide the agent's behavior
• SOPs help make the agent's behavior more predictable and debuggable for developers

Modular Prompting and Steering

• Strands introduces a "steering" feature that allows developers to provide external feedback and constraints to the agent during its execution
• This is done through hooks in the agent's lifecycle, where external systems can judge the agent's actions and provide feedback

Evaluations Library

• Strands has introduced an evaluations library to help developers build and run evaluations on their agents
• This includes the ability to generate synthetic data sets and integrate with online evaluation systems like Agent Core

Business Impact and Real-World Applications

Efficiency Gains in Agent Development

• The Strands team has seen significant efficiency gains in their own development process by using Strands agents to automate code generation and updates
• They report being able to save their principal engineer up to 4 hours of time on a single feature implementation

Diverse Use Cases

• Strands agents are being used by customers to build a wide range of applications, including chat experiences, data processing pipelines, and research assistants
• The model-driven approach and extensibility of Strands make it suitable for a variety of AI-powered applications

Ongoing Development and Community Involvement

• Strands is actively being developed, with new features like the TypeScript SDK and integrations with tools like Copilot Kit
• The Strands team encourages community involvement, with the project hosted on GitHub and opportunities to contribute model providers and other extensions