This tutorial builds a coding assistant that can answer a question, inspect files through declared tools, keep a resumable session, work inside a session-owned directory, and confine process tools. The final runtime is independent of its presentation, so you can keep the CLI or attach the TUI, Web UI, ACP, or MCP later.

What you will build

%%{init: {"theme":"dark"}}%%
flowchart LR
  U[CLI input] --> S[Harness Session]
  S --> L[Loop]
  L --> I[Inference Client]
  L --> T[Read-only Tools]
  S --> J[Session Store]
  S --> W[Workspace]
  T --> X[Sandboxed Process]
  S --> E[Events]
  E --> U

The tutorial uses one project named looprig-coding-assistant. Each page adds one boundary without changing the responsibilities added earlier.

How the pieces fit

Module	Responsibility in the assistant	Added when
Core and Inference	Messages, model identity, requests, responses, and streaming	First model call
LLM	OpenAI, Anthropic, Ollama, and other provider clients	Model configuration
Harness	Loop, Rig, Session, commands, events, tools, and shutdown	Agent runtime
Tools	Read, search, edit, process, interaction, and delegation definitions	Tool registration
Fsstore and Storage	Durable session history and workspace snapshots	Persistence
Sandbox	Native confinement for child processes	Process execution
TUI or Client	Terminal and framework-neutral browser presentation	Interface selection

You do not need to adopt the full stack at once. Inference is useful by itself. Harness becomes useful when you need a live agent runtime. Storage, Sandbox, protocols, workflows, and interfaces remain optional composition boundaries.

Tutorial path

Start with the project directory. You will have a working model call before introducing Harness.