What Is Agentic Coding? Definition, Tools and How It Works
Agentic coding is a way of building software in which an AI agent does the work end to end: it reads the codebase, plans the change, edits files across the project, runs commands and tests, checks the results and keeps going until the goal is met, with a developer setting the goal and reviewing the outcome. The developer moves from typing code to directing and reviewing an agent.
How agentic coding differs from autocomplete and chat
| Autocomplete | Chat assistant | Agentic coding | |
|---|---|---|---|
| Unit of work | The next line | An answer or a snippet | A whole task: a bug fix, a feature, a migration |
| Who edits files | You | You, copying from the chat | The agent |
| Runs commands and tests | No | No | Yes, within the permissions you set |
| Checks its own work | No | No | Yes: it runs tests, reads errors and tries again |
| Your job | Accept or reject suggestions | Ask and paste | Set the goal, review and approve |
How it works
An agentic coding session usually runs as a loop inside an agent harness, the software that gives the model tools, memory and permissions:
- Understand. The agent reads the relevant files, the project’s instructions (for example AGENTS.md or CLAUDE.md) and anything you point it at.
- Plan. It breaks the goal into steps. Many tools have a plan mode that shows the plan before anything changes.
- Act. It edits files and runs commands: installing dependencies, running the build, calling tools over MCP.
- Check. It runs the tests or another check that passes or fails, reads the output, and fixes what broke.
- Hand back. It summarizes what it did, often as a diff or a pull request for a person to review.
Large tasks are split further. An agent can hand parts of the work to subagents, each with its own context window, or several agents can run in parallel in separate git worktrees.
Tools for agentic coding
Most agentic coding tools are harnesses around a frontier model:
- Claude Code (Anthropic): “an agentic coding tool that reads your codebase, edits files, runs commands”, in the terminal, IDEs, a desktop app and the web.
- Codex (OpenAI): runs in a CLI, an IDE extension, the ChatGPT apps and the cloud, and “runs many agents in parallel”.
- Cursor: an editor with an agent, a CLI and cloud agents, with models from several vendors.
- OpenCode: “the open source coding agent”, with 200+ model providers.
- Google Antigravity: Google’s agentic development platform, with an app, a CLI and an IDE.
Side-by-sides with sources: Claude Code vs Codex and Cursor vs Claude Code.
What makes agentic coding work well
- A check the agent can run. Tests, a type check or a build that comes back pass or fail, so the agent can verify its own work instead of guessing.
- Short, specific project instructions. Commands, conventions and the rules the code doesn’t show, in AGENTS.md or CLAUDE.md.
- The right permissions. Let safe, repetitive commands run without asking, keep risky ones behind approval, and use a sandbox.
- Isolation for parallel work. One git worktree and branch per agent, merged through pull requests.
- Human review where it counts. Agents can produce a lot of code quickly; review stays the developer’s job.
Risks and limits
- Confident mistakes. Agents can produce plausible code that’s wrong in ways tests don’t catch, which is why review matters.
- Context limits. Very large codebases don’t fit in one context window; agents rely on search and on keeping the context focused.
- Security. An agent that runs commands can do damage if its permissions are too broad or if it follows instructions planted in files or web pages (prompt injection).
- Cost. Long sessions and many tool calls use a lot of tokens.
Agentic coding on a team
Each agentic coding session belongs to one developer. When several people’s agents work on the same codebase, they can collide: two agents change the same function, or one renames a field another still reads. Teams handle that with worktrees and pull requests, and with a shared place where the agents can see each other’s plans. elvex is that place: each person’s agent, Claude Code, Codex or Cursor, joins a shared thread, claims and hands off to-dos, and brings people only the decisions. See Claude Code multiplayer.
Agentic coding vs vibe coding
Vibe coding usually means building by prompting and accepting what the AI produces without reading it closely, often for prototypes. Agentic coding describes how the work is done, by an agent that plans, acts and checks, and is used for production code with tests and review.
- What is agentic coding?
Software development where an AI agent plans a change, edits files, runs commands and tests, and checks its own work toward a goal a developer sets, with the developer reviewing the result. - What are examples of agentic coding tools?
Claude Code, OpenAI Codex, Cursor’s agent, OpenCode and Google Antigravity. - How is agentic coding different from AI autocomplete?
Autocomplete suggests the next line for you to accept. An agentic coding tool takes a whole task, edits the files itself, runs the tests and fixes what fails. - Is agentic coding safe?
It can be, with the right setup: limited permissions, a sandbox for commands, a check the agent must pass, and human review before changes merge. - Can several developers use agentic coding on the same project?
Yes. Give each agent its own git worktree and branch, and give the agents a shared place to see each other’s plans, such as an elvex thread.