An agentic development environment (ADE) is an app for running AI coding agents. You give an agent a task, it plans the work, edits files and runs commands, and you review each change before it becomes part of your project. A code editor is built around a person typing code. An ADE is built around a person directing and checking agents, often several at once.
- Give a task
- Approve and review
- Tasks and worktrees. Each task kept apart
- Plans. Steps you can follow
- Permissions. What runs without asking
- Diffs and Git. Review, commit, push
- Codex
- Claude Code
- OpenCode
- Files
- Branches
What does ADE stand for?
ADE stands for agentic development environment. The name is new and still settling. Warp, a terminal app, says its product is "built as an agentic development environment", and Google calls Antigravity its "agentic development platform." Older software writing sometimes uses the same letters for "application development environment," which means something else.
How is an ADE different from an IDE?
An integrated development environment (IDE), such as VS Code or a JetBrains editor, is organized around files. You open a file, write code, run it and debug it. AI features in an IDE usually help with the code in front of you: they complete a line, explain a function or rewrite a block you selected.
An ADE is organized around tasks. A task names a goal ("fix the sign-up form"), the agent working on it, the branch it works in, and what it changed. The agent does most of the typing. You spend your time deciding what to ask for, answering the agent's questions, and reading what it did. Many people keep an editor open next to an ADE and use each for what it does well.
The two kinds of tool now overlap. Some editors add a separate view for running agents on whole tasks, and an editor built around an agent is sometimes called an agentic IDE. Some ADEs can also show and edit files. What separates them is where a tool expects you to spend your time: typing code, or giving tasks and reviewing what comes back.
How is an ADE different from a coding agent?
A coding agent is the program that does the work. It sends your request to an AI model, reads the model's answer, and carries out the steps the model asks for, such as opening a file or running the tests. Claude Code, Codex and OpenCode are coding agents. Anthropic says you can "steer and review" Claude Code from your desktop, a terminal, an IDE, Slack or the web, and OpenCode calls itself "the open source coding agent for any model".
An ADE is the workspace around one or more of those agents. It starts them, keeps their tasks apart, shows their plans and questions, holds their changes for review, and handles the Git steps afterwards. If you use one agent in a terminal for one task at a time, you may not need more. The extra structure helps once you run several tasks at once or want to hand a task from one agent to another.
What should you look for in an ADE?
- The agents you already use. Each agent has its own models, limits and subscription. An ADE that runs the agents you already pay for avoids a second bill for the same work.
- Separate worktrees. A Git worktree is a second copy of your project on its own branch. When each task has one, two agents can work at the same time without overwriting each other.
- Plans you can follow. A short list of steps, with the current one marked, tells you where a long task is without reading the whole transcript.
- Permissions and approvals. You should be able to choose what an agent may do on its own (read files, edit files, run commands) and get a clear prompt for anything beyond that.
- Review before commit. Changes should open as a diff with line numbers and signs, and committing, pushing and merging should each be a separate step that you take.
- Where it runs. Some tools run agents on your computer and some run them on the vendor's servers.
- Work that survives. Tasks should still be there after a restart, and you should be able to check on them away from your desk.
Should coding agents run on your computer or in the cloud?
When agents run on your own computer, they work in the same folders, tools and Git setup you use. Your code stays where it is, and the agent can run the same commands you would. The computer has to be on for the work to continue.
When agents run in the cloud, the vendor copies your repository to its own machine and works there. The task continues when your laptop is closed, and the result usually comes back as a branch or a pull request. In exchange, the vendor's machine needs access to your repository, and anything that only exists on your computer, such as uncommitted changes or local services, isn't there.
You can also use both: quick, hands-on tasks on your computer, and longer, well-defined ones in the cloud.
Where AI Integrator Code fits
AI Integrator Code is an ADE for Windows and macOS. It runs the coding agents installed on your computer: Codex, Claude Code, Cursor, Grok, Antigravity, Kimi Code, OpenCode, and other agents that support the Agent Client Protocol. Each one keeps its own sign-in and limits. If you don't have a subscription with an agent's maker, you can choose the AI Integrator runtime instead and run more than 40 models in Codex, Claude Code or OpenCode on your AI Integrator plan, as Code on your AI Integrator plan explains.
A task can move from one agent to another partway through. In the clip below, Codex fixes a broken sign-up form, and the task is then switched to Claude Code, which picks up from Codex's fix.
For the rest of the checklist above, Code has:
- Worktrees. A chat can run in its own worktree on a branch named after the chat. Merge brings its work into your project, and Undo reverses that merge.
- Plans. An agent working through a multi-step task posts its plan on one line at the top of the chat, with how many steps are done. The plan looks the same for every agent.
- Permissions. Each task has one of five levels: Read Only, Edit Files, Ask First, Auto and Bypass Permissions. On Auto, a second model reviews each step that reaches past the project's files, such as a command or a web request, and sends the ones it is unsure about to you.
- Review. Diffs show signs and line numbers as well as color. Commit and Push are separate buttons, and Code can open, review and merge pull requests.
- Subagents. A lead agent can hand focused jobs to helpers, up to six at once by default. Their work comes back to the lead chat, and a helper given its own worktree keeps its commits on a separate branch until you or the lead agent merge them. Subagents, plans and questions in Code shows how they work.
- Your phone. The iPhone app pairs with your computer, so you can answer an agent's request or check on a task away from your desk.
Code runs on your computer. Repositories, Code chats and agent sign-ins stay there, and AI Integrator does not run your code on its servers. On the AI Integrator runtime, the agent's requests to the model pass through your account, which doesn't store the requests or the replies. Chat and Agents can read your Code chats only if you turn that on for the computer. For work you want in the cloud, Code can start a Cursor Cloud Agent through an optional plugin, after you approve the repository, prompt and model.
Code does not replace an editor, so keep yours open alongside it. Quitting the desktop app stops the agents it started, and a scheduled task that was missed while the app was closed runs once when it opens again.


