What is Agentic coding?

Agentic coding is a way of building software in which an AI agent does a multi-step job on its own, instead of just suggesting the next line. You give it a goal, such as "add a password reset flow" or "fix this failing test," and it works through the task: reading the codebase, planning, editing files, running commands and checking the results.

How it differs from autocomplete

Early AI coding tools completed the line you were typing or answered a question in a chat box. An agentic tool acts. It uses tools such as a terminal, a file editor and a test runner, looks at what happened, and decides what to do next. That loop is what makes it an AI agent rather than a text generator.

The software that runs this loop, executes the tool calls and enforces limits is often called an agent harness.

How a typical run goes

1. You describe the goal, ideally with the constraints that matter.
2. The agent explores the project to find the relevant code.
3. It proposes a plan and makes changes across one or more files.
4. It runs tests or the program, reads the errors and tries again.
5. It hands back a result, often as a set of changes for you to review.

What it is good and bad at

It tends to work well on well-defined tasks with clear checks, like a bug with a failing test, which is the kind of problem measured by benchmarks such as SWE-bench. It struggles when the goal is vague or the important decisions were never written down. Giving the agent good information matters as much as the model, which is the focus of context engineering.

Why it matters

Agentic coding shifts a developer's time from typing code toward specifying work and reviewing output. Reviewing generated changes carefully becomes the core skill, because an agent can produce plausible code that is wrong. It is related to vibe coding, but the two are not the same: vibe coding means accepting code without reading it closely, while agentic coding describes the agent's multi-step way of working, and a careful team can pair it with strict review.

Safety basics

An agent that can run commands can also do damage. Run it in an agent sandbox, give it only the access the task needs, and keep a log of what it did.

An example

You ask an agent to upgrade a library. It finds every place the old version is used, changes them, runs the test suite, fixes two failures it caused, and returns a summary for you to check before anything is merged.