Course home

Architecting a safe agentic system

A multi-page, learn-by-building course. Each lesson teaches one architecture component, then you wire it into the next — until you have a gated, observable Instagram content agent you understand end to end.

~5 weeks at a steady pace Local-first · LM Studio ready Safety before Instagram

How to use this course

  • Read one page at a time. Build the checkpoint before moving on.
  • After every step, answer the four notes questions at the bottom.
  • Do not skip policy, runs, or evals to “get to publishing faster.”
  • Open these files in a browser (or serve docs/ with any static server).

Start paths

1. Core principles

Cognition vs authority, budgets, HITL, harness-agnostic design.

2. Architecture map

The target system diagram and what each layer owns.

3. Begin Phase 0

Sketch the system against the current repo, then start Step 01.

4. Full syllabus

All 18 lessons with week groupings and definition of done.

What you will build toward

Harness (MCP) + React (IG + monitor + approve) ↓ Interface (MCP + REST + SSE) ↓ Orchestrator (state, policy, runs) ↓ Agent loop ↔ Model adapter (LM Studio / OpenAI-compatible) ↓ Skills/tools → Policy gate → Side effects (media, IG) ↓ Memory (brand/learn) + Evals
Golden rule

The model proposes. The orchestrator authorizes. Humans approve irreversible actions. If the LLM can publish to Instagram without a gate, you do not have a safe agentic system.