Use an existing harness

This path explores the same capabilities as labs/, but configures a vendor harness instead of building its loop. Choose Codex CLI, Claude Code or GitHub Copilot CLI. Codex SDK has a separate directory for future exercises.

Each CLI lab is an independent exercise across the complete capability sequence, from setup and baseline to comparison, with corresponding standalone folders for the other CLIs. These are not yet comparable scored runs: vendor CLIs own their loop and may expose built-in tools even in read-only mode. Record which capabilities are built in, configurable or unavailable instead of calling the baseline a bare model call. Advanced workflow exercises identify human-orchestrated steps rather than claiming a native graph runtime.

For Foundry, prefer Entra ID where the chosen harness and endpoint support it. The Codex CLI guide currently says Entra ID is not supported for Codex; its exercises use a learner-provided resource API key. The Claude Code guide documents both Entra and key authentication. Copilot CLI’s documented Azure BYOK example uses an API key; do not confuse the Foundry resource management skill with inference authentication. Scoped instruction rules appear in CLI Labs 2B and 5; Lab 6 contrasts them with tool permissions and approval. Never commit keys or access tokens; the build-your-own labs/ path retains its Entra-only contract.