05. MCP vs APIs vs Function Calling vs Agent Skills: First Principles

Place four complementary mechanisms in one request path and identify their separate responsibilities. Episode 5 of 23 | AI Agents for Developers: From First Principles to Production Place four complementary mechanisms in one request path and identify their separate responsibilities. Original researched, welcome-first lesson with worked examples and explicit limits. Local Python results use scripted model decisions; architecture illustrations are not live-model benchmarks. Original script, visuals and examples. Narration generated locally with the speaker's authorization. Educational scenarios distinguish measured Python outcomes from illustrative architecture and scripted model choices. Learning references: Anthropic: https://www.anthropic.com/engineering... OpenAI: https://developers.openai.com/api/doc... Model Context Protocol: https://modelcontextprotocol.io/docs/... Agent Skills: https://agentskills.io/home Anthropic: https://www.anthropic.com/engineering... OpenAI: https://developers.openai.com/api/doc... Anthropic: https://www.anthropic.com/engineering... AWS Builders' Library: https://aws.amazon.com/builders-libra... Chapters 0:00 Start with the missing responsibility 1:03 A software capability exists before the model 2:18 An API is a contract, not just an HTTP endpoint 3:34 Function calling is how a model proposes an operation 4:46 Follow a direct tool connection before adding MCP 6:04 MCP separates the host, client and capability provider 7:22 Discovery and invocation are different protocol steps 8:49 Local and remote change deployment, not the need for trust 10:03 A Skill supplies the reusable way of working 11:20 Assemble the full path without confusing the layers 12:38 Use the recorded local run to verify what actually exists 13:55 Choose a layer only when it addresses a real need 15:19 Remember the responsibilities rather than the acronym contest