Agentic Engineering

Agentic engineering is the practice of building software where AI agents do most of the typing and humans do most of the thinking. The agent reads the codebase, proposes a change, opens a PR, runs the tests, and waits for review. The human reviews intent, edits as needed, merges. This is the workflow that emerged across teams using Claude Code, Cursor agents, Devin, and similar tools through 2025 and 2026. Karpathy distinguished it from vibe coding explicitly: vibe coding is fine for throwaway work; agentic engineering is the discipline for production code that will outlive the immediate session.

The shift from "AI as autocomplete" to "AI as a contributor with a PR" is the practical change behind agentic engineering. It requires three things in your codebase. First, fast and reliable tests that an agent can run unattended (slow tests cause agents to skip them). Second, clear architectural conventions written down somewhere the agent can read (otherwise every PR drifts in style). Third, code review discipline applied to AI PRs the same as human PRs (the failure mode is rubber-stamping AI work that compiles but encodes wrong assumptions). Teams that get these three right ship 2 to 4x more code than they did pre-agent.

Adopt agentic engineering as your default workflow for any code that ships to users. The setup cost (fast tests, written conventions, review discipline) pays back within the first few weeks.

• Anthropic: Claude Code overview
• Karpathy on agentic vs vibe coding

Related Terms