This 4-hour workshop focuses on agentic AI engineering. The instructors, Jon Krohn and Ed, cover defining AI agents, design principles for building agentic systems, and hands-on coding exercises using OpenAI Agents SDK, Crew AI, and Anthropic's MCP. The workshop aims to equip participants with the skills to build and deploy agentic AI systems.
Definition of AI Agents: AI agents are programs where large language model (LLM) outputs control the workflow, offering autonomy in problem-solving. This involves multiple LLM calls, tool usage, environmental interaction, a planner, and crucial autonomy.
Agentic AI Opportunity: There's significant current business value in building AI applications using agentic workflows, particularly for repetitive tasks. This can lead to an AI ROI flywheel, where initial success fuels further investment.
Agentic AI Frameworks: Several frameworks simplify agentic system development. These range from simple (no framework, MCP protocol) to more complex (OpenAI Agents SDK, Crew AI, LangChain/LangGraph, Microsoft Autogen). The workshop focuses on OpenAI Agents SDK and Crew AI.
OpenAI Agents SDK: A lightweight, flexible framework for creating agents, simplifying tool integration compared to directly using OpenAI's chat completions API. It offers built-in guardrails for safe deployment.
Crew AI: A framework tailored for multi-agent systems, allowing configuration via YAML files. It simplifies complex agent interactions and provides features like code execution within Docker containers.
Anthropic's MCP (Model Context Protocol): A standard protocol for connecting agents to data sources and tools, eliminating the need for glue code and enabling easy integration of various components.
Building an Autonomous Trading System: The workshop culminates in building an autonomous trading system using multiple agents, various MCP servers, and real-time (or delayed) market data. This demonstrates the practical application of agentic AI.
The three steps are: 1) Create an instance of an agent object; 2) Use a context manager called trace to describe the action; and 3) Call runner.run to initiate the agent workflow.
The five workflow design patterns are: 1) Prompt chaining; 2) Routing; 3) Parallelization; 4) Orchestrator-Worker; and 5) Evaluator-Optimizer. These workflows involve predefined, often sequential, steps, with varying degrees of LLM and code involvement. Agents proper, in contrast, are open-ended systems where LLMs dynamically direct their processes and tool usage, exhibiting greater autonomy and potentially more complexity and risk.
The five steps are: 1) Create a project using crewai create [project_name]; 2) Edit the agents.yaml and tasks.yaml configuration files; 3) Write code in crew.py to stitch together agents and tasks; 4) Finalize setup in main.py, including specifying inputs; and 5) Run the project using crewai run.
The three key components are: 1) The host (an application like a desktop app or custom agent architecture); 2) The MCP client (residing within the host and connecting one-to-one with an MCP server); and 3) The MCP server (providing tools, context, and prompts, which can run locally or remotely).
(This is open-ended and requires a user-specified question based on the transcript.)
