Level 2, intermediate. Your first weeks with Claude Code or Codex: how an agent uses tools and files, when to split work across sub-agents, and how to keep it on the rails.
Updated July 19, 2026
This Level 2 guide is for people who already chat comfortably with AI and are trying a coding agent like Claude Code or Codex for the first time. You will learn how agents use files, tools, context, sub-agents, skills, MCP connectors, and verification loops so you can let an agent do real work without losing control of your project.
Who this is for: You are trying Claude Code or Codex for the first time and want your agent to use tools, edit files, and delegate to sub-agents well.
A coding agent is not just a chat window that knows programming; it is a task runner with project access, tools, checks, and a workflow you supervise.
Read the deep diveTreat `AGENTS.md` or `CLAUDE.md` as a project map, not a storage locker: give the agent enough direction to find the right context at the right time.
Read the deep diveA coding agent works best when you give it a small, inspectable job with known inputs, real constraints, planned checkpoints, and an unambiguous finish line.
Read the deep diveGive your coding agent more reach one permission at a time, so every new tool expands what it can do without hiding what it might break.
Read the deep diveSub-agents are useful when you treat them like narrow workers with clear inputs, limits, and review gates, not like extra chat windows you hope will sort themselves out.
Read the deep diveAn agent loop is a small repeatable system: start it with a trigger, let the agent act through tools, verify the result, stop on a clear condition, and leave a memory log for the next run.
Read the deep dive