The category of AI systems that don't just autocomplete code but take goals and execute multi-step engineering tasks — reading files, running commands, editing code, and verifying results. Andrej Karpathy framed them as "the first real agents that work" because code has fast, cheap feedback loops (compilers, tests) that ground the model. Examples include Claude Code, Codex, Cursor's agent mode, and Devin. For founders, code agents are reshaping what a small team can ship — the leverage per engineer is going up faster than any tooling shift in a decade.