What it is
Agentic coding means an AI agent takes a task, writes the code itself, runs the tests, fixes the errors and opens a PR. The engineer's role shifts from typing lines to framing tasks, reviewing and owning architecture. Claude Code, Cursor Agent, Devin, GitHub Copilot agent mode.
Where it came from
Autocomplete (Copilot, 2021) → chat assistant (2023) → autonomous agent (2024–25) that works a task through on its own, across many steps. Anthropic, OpenAI and GitHub all built "agent mode" into their dev tools.
Why it took off
Anthropic shipped and opened up the Agent Skills standard for enterprise workflows (code, legal, finance, data); the skills repository passed 20k stars on GitHub. Enterprises are running them in production. The reason: an agent doesn't close a suggestion, it closes a complete unit of work — a feature, a fix, a refactor.
How to apply it now
- Give the agent well-bounded tasks (a bug, a feature, a refactor), not "build the whole project".
- Always review the output: an agent is confidently wrong, especially on edge cases and security.
- Use reusable skills and prompts as your team's "procedures".
- Measure throughput (tasks closed), not lines of code.
What to watch out for
An autonomous agent with access to your repo and CI is also a risk: it can delete things, break things, leak secrets. You need isolation, branches and a check before merge. Audit skills pulled from other people's repos (supply chain). Generated code is still your responsibility.