🗓️ 19092026 2037
AI applications become useful when a language model can use context and act through controlled capabilities. This collection explains the layers around that idea, not how models are trained.
Start here
An AI application has four separate jobs:
- Model — turns a request and context into a response or a proposed tool call.
- Application — supplies context, runs tools, and enforces permissions.
- Workflow — decides whether one pass is enough or work needs multiple steps.
- Integration — connects the application to external systems such as files, databases, and APIs.
The model can propose an action. The application must decide whether to allow and execute it. llm_tool_use explains this boundary.
Two learning paths
Build an agent
- Start with llm_tool_use.
- Read agentic_design_patterns to choose a workflow.
- Read agent_memory_and_state only when work must survive more than one model call.
- Use claude_agent_sdk when an SDK fits the workflow, or building_with_claude to compare implementation choices.
Connect external systems
- Start with the MCP overview.
- Read mcp_architecture and mcp_transports.
- Read mcp_authorization only when a remote server needs user-authorized access.
- Read custom_mcp_servers when an existing server cannot safely expose your system.
How the ideas fit
user request
|
application / agent workflow
|-- model uses [[llm_tool_use]]
|-- optional memory and review loops
`-- optional MCP client
`-- MCP server
`-- database, files, API, or service
MCP is an integration standard, not an agent framework. It tells an application how to discover and communicate with a capability provider. It does not decide what the model should do or grant permission for an action.