A Cursor AI tutorial that takes one failing test to a reviewed commit through Plan Mode, the Run Mode that decides whether tests run unasked, and Agent Review.
What an autonomous AI agent can do without a human, mapped to the permission modes, sandboxes and hard caps that Claude Code, Codex and Copilot document.
What an MCP server is, a minimal Python server you can run, the Claude Code command to register it, and the failure modes that make a first server go silent.
Test AI-generated code across contracts, boundaries, permissions, migrations, retries, concurrency, and rollback with a reusable failure matrix.
Use an AI coding workflow that turns issue contracts, tests, diffs, review findings, and release checks into evidence instead of confident summaries.
Use an AGENTS.md file to give coding agents scoped repository rules, nested overrides, exact verification commands, and deployment anti-drift boundaries.
Follow a Claude Code tutorial that inspects, plans, edits, and verifies one small repository change with exact checkpoints and test evidence.
Compare Codex vs Claude Code with the same five repository tasks, evidence checklist, permission boundaries, and a decision method you can reproduce.
What AI coding agents actually are, how they work, what they genuinely nail, and where they confidently break, from a dev who ships real apps with them.