[Workshop] AI Pipelines and Agents in Pure TypeScript with Mastra.ai — Nick Nisi, Zack Proser

Key Concepts

• AI Pipelines: Composable and typed sequences of steps for AI tasks.
• Workflows: Discrete sets of steps to accomplish a task, allowing data transfer and error handling.
• Tools: Functions that agents can call for tasks like file access, API calls, or database queries.
• Agents: Interfaces (often chatbots) that interact with humans using natural language to trigger workflows.
• MCP (Model Context Protocol): An open standard for universal plugins or tools that AI agents can call.
• Zod: A TypeScript library for schema validation and coercion, ensuring reliable output from LLMs.
• ImageFlip: An API used for finding base memes and creating new captions.
• Mastra: A framework for building production AI applications in an agentic manner.

Introduction and Concepts (20 minutes)

• The workshop focuses on building an AI-powered meme generator using TypeScript and the Mastra framework.
• The patterns used are applicable to production AI applications.
• The format includes introductions, concept explanations, collaborative coding, and a wrap-up with Q&A.
• Workshop materials are available in the AI Engineer Slack channel under "workshop AI agents with Mastra".
• Work OS is a company that builds developer tools for enterprise features like SSO, SAML, directory sync, audit logs, and fine-grained off.
• Work OS is expanding into AI, focusing on securing AI and attaching identity to agents.
• Mastra is a framework for building production AI applications in an agentic manner.
• Key components of Mastra include workflows, tools, and agents.
• Mastra offers built-in persistence, memory, observability, and evaluation.

Workflows: Composable Pipelines

• Workflows are composable pipelines of discrete steps.
• They allow chaining multiple steps together to complete a task.
• Data can be passed between steps.
• Zod is used to validate inputs and coerce outputs, ensuring reliability.
• Workflows provide a deterministic set of steps for a task.
• The API is similar to RxJS, starting with one thing and then mapping to something else.

Tools: Functions for Agents

• Tools are functions that agents can call.
• They provide access to file systems, APIs, databases, and custom business logic.
• Tools give LLMs the ability to act on behalf of the user and access data.
• In the workshop, the agent calls a workflow rather than individual tools.

Agents: Ideal Human Interfaces

• Agents are considered ideal human interfaces, often in the form of chatbots.
• They allow users to describe their needs in natural language.
• The system can then use workflows to complete the task.
• Agents combine prompts with workflows or tools to perform specialized tasks.
• Agents can introspect the state and make decisions.

MCP (Model Context Protocol)

• MCP is an open standard for bringing universal plugins or tools to AI agents.
• It was developed by Anthropic and is used in Claude, Cursor, Windsurf, and OpenAI.
• MCP provides an easy-to-use API for tools to interact with AI agents.
• Example: Using the GitHub MCP server in Claude to access issue context and suggest solutions.
• mcp.shop is a demo where you can order a shirt using MCP.
• MCP Night was held at the Exploratorium, indicating growing interest in the protocol.

AI Meme Generator: Workshop Project

• The workshop project is an AI meme generator.
• The input is a user's frustration.
• The workflow finds the best base meme, edits captions, and publishes a new meme.
• The project uses OpenAI for understanding and ImageFlip for image manipulation.
• The patterns used are production-ready.

Workshop Structure and Setup

• The workshop is self-paced.
• The workshop.md file in the repo contains the entire course.
• Git branch checkpoints are provided for each phase.
• Assistance is available from the presenters and the Mastra core team.
• Understanding is prioritized over completion.
• The workshop is a snapshot in time due to the rapidly changing nature of AI tools.
• Step zero branch provides the baseline to start from.
• The goal of the first step is to see the Mastra playground.

Step 1: Creating a Basic Workflow

• The goal is to create a basic workflow with a single step.
• The workflow will extract the type of frustration from user input.
• The Mastra playground allows testing the workflow.
• The playground shows agents, networks, tools, MCP servers, workflows, and runtime context.
• The playground provides a URL (localhost:411) to access the interface.

Step 2: Building the Full Workflow

• The goal is to create all workflow steps and chain them together with data mapping.
• The steps include extracting frustration, finding a base meme, generating captions, and generating the meme.
• The workflow will generate a meme and upload it to ImageFlip.
• The Mastra playground allows testing the full workflow.
• ImageFlip requires a username and password for a higher rate limit.

Step 3: Integrating with an Agent

• The goal is to create an agent that can call the workflow.
• The agent is given a prompt and knowledge of the workflow.
• The agent provides a chat interface for users to interact with the workflow.
• The Mastra playground allows testing the agent.
• The agent can provide human feedback and show the image inline.

Key Takeaways and Next Steps

• The workshop covered workflows, agents, type safety, error handling, and structured generation.
• Mastra provides benefits such as easy debugging and rapid iteration.
• Next steps include exploring tool creation, adding a vector database, and implementing RAG.
• mcp.shop is a demo to order a t-shirt using MCP.

Q&A and Discussion

• Discussion on the use of Vim and AI integrations.
• Discussion on the use of cursor rules and MCP doc server.
• Discussion on the use of durable execution for agents.
• Discussion on deployment platforms and strategies.
• Discussion on schema validation and error handling.
• Discussion on the use of DAGs and Airflow.
• Discussion on caching with Mastra.
• Discussion on the use of agents in production.
• Discussion on the potential for non-technical users to use AI tools.
• The Mastra team thanked the participants for attending the workshop.
• The presenters thanked the Mastra team for creating the framework.