You will first learn Python and ordinary application engineering. Those skills let you understand the control loop instead of treating a framework as magic. Next come model inputs and outputs, retrieval, evaluation, and access control. Only then do the agent-specific modules add planning, memory, coordination, and durable execution.
Use fake model decisions and sandbox tools while learning runtime behavior. This makes invalid arguments, timeouts, retries, and rejected approvals reproducible without API costs. Later, evaluate actual model decisions separately. Passing fake-client tests does not establish real model quality.
Keep one evolving support-agent project. Start with a fixed workflow, add a bounded tool loop, then persistent state and approval. Each added capability should have a test and an explanation of the problem it solves. All stages are self-paced.
Exercise: write a three-step workflow on paper, then manually simulate a missing document and a rejected action. Record the state after each step.
Check: you can say why the workflow stops and what, if anything, it changed before stopping.