Day 4 · Week 1
Review, repair, and practise
Scheduled session · Thursday, 17 September 2026
Our planned session
We design the agent loop and its safety boundary: goal → model decision → allowlisted tool request → argument validation → tool result → final answer. We define what requires human approval before writing the implementation.
Weekly outcome
A working AI agent powered entirely by a local Ollama model.
Readiness check
We need an editor and one available assistant; Day 1 includes a browser fallback for JavaScript.
Our 20-minute practice
We spend 20 minutes on one small JavaScript/TypeScript problem: first attempt, then agent critique and one test.
Human approval gate
We can explain the difference between a prompted chatbot and an agent, then demonstrate that only an allowlisted tool with validated arguments can run.
This chapter has not been published yet.
We plan to publish each chapter one day before its scheduled session. Session results record work we actually performed and reviewed. Neither the calendar date nor publication alone marks this activity complete.
Expected file: content/365-days-to-fullstack-ai-engineer/day-004.md