Kyle Stratis

Book Author: AI Agents with MCP

• Kyle Stratis
  • Homepage Very intersting!!
  • LinkedIn
  • Book
    • AI Agents with MCP
• slides

NoteNotes

• Event context

  • The transcript is from a virtual book signing with Kyle Stratus, founder/principal consultant of Stratus Data Labs and VP of Engineering at the SQL Institute.
  • The discussion centers on his book AI Agents with MCP, currently in early access on O’Reilly and available for pre-order.

• Motivation for the book

  • The book grew out of Kyle’s difficulty building his first Model Context Protocol (MCP) client in early 2025.
  • At the time, most MCP documentation focused on servers, while client-side implementation guidance was sparse.
  • His goal is to provide an authoritative, practical reference for building with MCP in Python, covering clients, servers, security, transports, architecture, and deployment.

• What MCP solves

  • MCP decouples tools, prompts, data, and resources from the agent application itself.
  • Without MCP, tools are often tightly coupled to a specific agent framework or application.
  • With MCP, service providers can expose agent-ready capabilities once, and application developers can reuse them across agents and use cases.
  • This makes MCP especially useful for enterprises and platform teams that need standardized, distributable agent interfaces.

• Advice for building AI agents

  • Standard software engineering principles still apply: clean architecture, code quality, and avoiding unnecessary duplication.
  • Agent systems introduce a broader security attack surface because models can interpret arbitrary natural-language inputs.
  • Kyle highlights Simon Willison’s “lethal trifecta”: privileged data access, external-world access, and code execution. Combining all three in one agent can create serious data-exfiltration risks.
  • A safer architecture separates these capabilities across agents, sub-agents, or coordinated components.

• Evaluation and observability

  • Agent evaluation should happen both during development and in production.
  • Tools like Promptfoo can help test tool choice, prompt behavior, and expected user scenarios.
  • Observability and tracing are necessary to understand what users actually ask, how the agent responds, and where it fails.

• Python CLI discussion

  • The workshop code uses the standard Python pattern if __name__ == "__main__": to run code directly from the command line.
  • Basic CLI interaction can be handled with input() and print().
  • For richer command-line interfaces, Kyle mentions libraries such as Click, Typer, and Argparse, with a preference for purpose-built CLI packages over bare Argparse.

• Skills versus MCP

  • Skills are described as selectively retrievable prompts that models can use for specific tasks.
  • They are useful for personal workflows, small-team automation, code review, editing, drafting, note expansion, and lightweight repeated tasks.
  • Their limitations include weak versioning, difficulty pinning dependencies, inconsistent model usage, and often being tied to a specific model ecosystem.
  • MCP can also expose prompts, but it additionally supports tools, resources, data access, sampling, authentication, and richer interaction patterns.

• When to use a skill

  • Use a skill when the task is mostly prompt-based, lightweight, personal, or limited to a small team.
  • Example: using Claude to generate language-learning flashcards from notes.
  • Skills are lower-overhead and easier to create than a full MCP server.

• When to use an MCP server

  • Use an MCP server when the capability resembles an external API, needs authentication, must be distributed broadly, requires versioning, or exposes tools/data rather than just a prompt.
  • MCP is more appropriate for enterprise distribution, service-provider integrations, and advanced interaction patterns such as elicitation.

• Overall takeaway

  • MCP is best understood as infrastructure for distributing reusable, agent-ready capabilities.
  • Skills are excellent for lightweight prompt augmentation; MCP is better for robust, scalable, authenticated, and reusable agent integrations.
  • The practical challenge in modern agent development is not only making agents useful, but making them secure, observable, evaluable, and maintainable.