Explainer

Agentic coding, explained in plain English

Updated November 17, 20269 min read

Agentic coding is a way of building software where you give an AI a goal and it carries out the work itself: it reads your project, plans, edits files, runs commands, checks the results and fixes its own mistakes. You act more like a manager who sets the task and reviews the outcome. Tools such as Claude Code, Codex, Google Antigravity and GitHub Copilot's agent all work this way.

What is agentic coding? A plain definition

"Agentic" comes from "agent": something that acts on your behalf. In agentic coding, the AI is given tools it can use without you copying and pasting. It can open any file in your project, write new ones, run the app, run tests and sometimes search the web. It uses those tools again and again, in a loop, until the goal you gave it is met or it gets stuck and asks you.

Compare that with the first wave of AI coding help. Early tools finished the line you were typing, or answered a question in a chat window. You still had to decide where the code went, paste it in, run it and spot the error. With an agent, all of those chores move to the AI. You describe the outcome ("users should be able to reset their password by email") and judge whether it works.

A simple picture: autocomplete is a calculator, a chat assistant is a helpful friend on the phone, and an agent is a builder you have let into the house with the keys. That last one gets far more done, and it also means you need some house rules.

Words you will meet when reading about agents

Agent
An AI program that can take actions (read files, run commands) and keeps going until a task is done.
Model
The AI brain inside the agent, such as Claude, GPT or Gemini. One agent can often switch between models.
Tool call
A single action the agent takes, like "open this file" or "run the tests".
Context window
How much text the model can hold in mind at once. Long sessions fill it, and the agent starts forgetting early details.
Permission prompt
The moment an agent stops and asks "may I run this command?" before doing something.
Plan mode
A setting where the agent studies your project and writes a plan, but changes nothing until you approve.
Instruction file
A note in your project the agent reads first, with your rules. Claude Code reads CLAUDE.md, Codex reads AGENTS.md, Gemini CLI reads GEMINI.md.
Subagent
A helper agent that the main agent sends off to do one part of the job, so its own memory stays clear.
Worktree
A second copy of your project folder that git keeps linked to the first, so two agents can work without bumping into each other.
Pull request
A proposed set of changes that a person reviews before they become part of the project.

How an agent loop works, one turn at a time

Every agentic coding tool runs some version of the same cycle. Picture a request like "add a dark mode switch to the settings page".

  1. 1

    Read

    The agent looks around. It lists folders, opens the settings page file, finds where colours are defined and reads your instruction file. Nothing changes yet.

  2. 2

    Plan

    It writes down what it intends: add a toggle to the settings page, store the choice, swap the colour values when dark mode is on. In tools with plan mode, you can read and edit this plan before any file is touched.

  3. 3

    Act

    It edits the files and may run commands, for example installing a small package or starting the app. Depending on your settings it asks before each step or just carries on.

  4. 4

    Check

    It reads what came back. Did the app start? Did a test fail? Did the page throw an error? Some agents can open a browser, click the new switch and look at the result.

  5. 5

    Repeat

    If something broke, it goes back to Read with the new error in hand and tries again. When the checks pass, it stops and writes you a summary of what it changed.

The quality of the loop depends heavily on the Check step. An agent that can run real tests, or see the actual page, catches far more of its own mistakes than one that only reads its code and hopes. That is why good projects give the agent a quick way to test, and why Google Antigravity lets its agents open a browser and record what they see.

On a big job the loop can run dozens of times. Each turn uses some of your plan's allowance, which is why long agent sessions use up usage limits much faster than chatting does.

From autocomplete to agents: a short history

Agentic coding grew quickly, but the big steps have clear dates. These are the ones we could confirm from the companies' own announcements:

  • June 29, 2021: GitHub launches a technical preview of Copilot, an "AI pair programmer" that suggests code as you type. This is the autocomplete era.
  • March 12, 2024: Cognition introduces Devin and describes it as a teammate that can complete tasks on its own for an engineer to review.
  • February 2, 2025: Andrej Karpathy coins "vibe coding" in a post on X.
  • February 24, 2025: Anthropic releases Claude Code as a limited research preview, a command line tool for agentic coding.
  • April 16, 2025: OpenAI releases Codex CLI, an open source agent for the terminal.
  • May 19, 2025: GitHub announces the Copilot coding agent, which takes an issue and opens a draft pull request.
  • June 25, 2025: Google releases Gemini CLI, an open source terminal agent.
  • November 2025: Google launches Antigravity, an agent-first editor, alongside Gemini 3.
  • June 18, 2026: Google moves free and Google AI Pro and Ultra users from Gemini CLI to the new Antigravity CLI.
“There's a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.”
Andrej Karpathy, post on X, 2 February 2025

Karpathy's post named a habit. The agents released in the months around it are what made that habit practical for people who had never written a program.

Agentic coding tools you can install today

Here are the main agentic coding tools as of November 2026, sorted by where they run. Prices and limits change often; our ranking of the best AI coding agents keeps the current figures.

  • In the terminal: Claude Code, Codex CLI, Antigravity CLI, GitHub Copilot CLI, and open source options like OpenCode and Aider.
  • In an editor: Cursor's Agent, Copilot agent mode in VS Code, the Antigravity IDE and Devin Desktop (formerly Windsurf).
  • In the cloud: Codex on the web, the Copilot coding agent, Google's Jules and Cognition's Devin. These work on a remote machine and send back a pull request.
  • In the browser for non-coders: Lovable, Bolt, Replit and v0, which build and host whole apps from a chat.

Agentic coding vs vibe coding

People mix these up all the time. Agentic coding describes what the tool does. Vibe coding describes how the person works. You can do either without the other, though they usually go together.

Agentic codingVibe coding
What it namesA kind of AI tool that acts on its ownA style of working with AI
Who reads the codeOften a developer reviews itOften nobody does
Main skill neededSetting clear goals and reviewing changesDescribing what you want and testing the result
Typical userDevelopers and teamsFounders, designers, hobbyists, anyone
Typical toolsClaude Code, Codex, Copilot coding agentLovable, Bolt, and the same agents with less review
Biggest riskAgent makes broad changes you didn't expectHidden security holes nobody looked for

A developer using Claude Code who reads every change is doing agentic coding but not vibe coding. A beginner accepting every change from the same tool without reading it is doing both. For the full story on the second style, read what is vibe coding.

Agentic software development in teams

Agentic software development is the team version: agents are given real tickets, work in parallel, and their output goes through the same review as a person's would. A typical setup in 2026 looks like this. One agent builds a feature, another writes tests for it, a third fixes a bug report, and a human reviews each pull request before it is merged.

Running several agents on one project needs a way to keep them apart, or they overwrite each other's edits. The standard answer is git worktrees, explained in our worktrees guide. Inside one agent you can also split work up with subagents, each with its own clean memory.

Keeping track of many agents is its own job. Desktop cockpits such as VibeCodingDeck exist for that: it runs up to 12 agents in one grid on Windows or an Apple-silicon Mac, gives each pane a status light, and puts each agent on its own safe copy of the project so you merge back only the work you like.

What can go wrong when an AI agent drives

An agent with the keys can do a lot of good quickly. It can also do damage quickly. These are the problems people actually run into.

Watch out

Wide changes you didn't ask for. Asked to fix one button, an agent may tidy up ten other files on the way. Take a git commit or a save point before every big task so you can roll back in one step.

Watch out

Commands that delete or overwrite. Agents can run any command you allow. Keep permission prompts on for real projects. Claude Code's bypass permissions mode is meant for sandboxes, such as a container or a throwaway virtual machine, not your main computer.

Watch out

Secrets in the wrong place. An agent may paste an API key straight into code and then commit it. Keep keys in environment variables (settings stored outside your code) and tell the agent so in your instruction file.

Watch out

Code that works but isn't safe. Agents are good at making things run, less good at locking them down. Logins, payments and user data need a careful review. Our vibe coding security guide lists what to check.

Note

Runaway usage. A loop that keeps failing keeps spending your plan. If an agent has tried the same fix three times, stop it, explain the problem differently, or hand the task to a different agent.

Agentic coding best practices checklist

None of these need technical skill. They are habits, and each one saves you from a specific kind of bad afternoon.

  • Write a short instruction file (CLAUDE.md, AGENTS.md or GEMINI.md) with your rules: what the app is, which tools it uses, what never to touch.
  • Start big or risky tasks in plan mode and read the plan before saying yes.
  • Give one clear goal per request. "Add search to the products page" beats "improve the site".
  • Tell the agent how to check its work: which command runs the tests, or which page to open.
  • Commit or take a save point before each task, so undo is one step.
  • Keep permission prompts on outside of a sandbox.
  • Keep API keys and passwords out of code, in environment variables.
  • Clear or restart the chat when the context window is nearly full, rather than letting the agent forget halfway.
  • Give parallel agents their own worktree so they never edit the same files at once.
  • Read the summary and test the result yourself before you ship. Ask a second agent to review anything important.

Once these are second nature, try running two agents at once on separate tasks. That is where agentic coding starts to feel like having a small team rather than a single helper.

Sources

Run up to 12 AI coding agents in one window.

Claude Code, Codex, Gemini, Grok and more, side by side. Windows and Mac. From $29 a month.

Get started

Keep reading