It's a narrower, more autonomous category than a general coding assistant that only suggests completions: a software engineering agent is expected to plan its own steps, use a terminal and version control, and close the loop against evals like SWE-bench without a human approving every intermediate edit. Andrew Ng's August 2026 AI Engineering Skills Map lists 'using coding agents' as a distinct, learnable skill precisely because effectively directing one requires knowing how much autonomy to grant versus when to intervene.