Written by the team developing Dure, a workspace for coding agents. Guides describe behavior documented in the linked product sources on the date shown, not hands-on benchmark results.
An agent development environment (ADE) is the workspace around a coding agent, not the agent itself. This page defines the term as current product documentation uses it, separates it from two unrelated meanings of the same acronym, and lists what such a workspace actually owns. Every claim below links to the product source it came from, checked on the date above.
What does the agent do, and what does the environment do?
A coding agent is a program that changes your code on its own instructions rather than waiting for each keystroke. Anthropic describes Claude Code as "an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools." Claude Code overview
These agents arrive as command-line tools and editor extensions. OpenAI lists Codex on a CLI, an IDE extension, a cloud service, and the ChatGPT desktop, mobile and web apps. Codex documentation
The environment is the other half of the job: deciding where each agent runs, keeping one agent's files out of the next agent's way, holding the work when you close the window, and showing you the result in a form you can answer. One agent on one task needs none of this written down, because you are holding it in your head. The question appears when several tasks run at once.
Why do the same three letters mean three different things?
The acronym is still settling, and three current uses do not describe the same product.
| What "ADE" refers to | The definition its source gives | Source |
|---|---|---|
| A workspace for coding agents | "A development tool — typically a terminal, IDE, or CLI — built around AI coding agents as first-class participants rather than as an add-on." | Warp, 2026-08-31 |
| A platform for orchestrating agents | "A developer platform built natively for AI agent orchestration, multi-threading, and human-agent collaboration." | Augment Code |
| A toolkit for stateful agents | A "comprehensive toolkit for creating, testing, and monitoring stateful agents," built around persistent memory rather than code changes. | Letta |
The first two describe the same category in different words; the third is a different product category that happens to share the letters. Letta's ADE is for agents that carry memory between conversations, and its documentation never mentions coding agents. If a search result mixes these together, check which one the page is actually about.
Both spellings are also in use. Warp and Augment Code write "agentic development environment," while Orca's site and Dure's documentation write "agent development environment." Orca · Dure introduction
What does an ADE own?
Isolation: one checkout per task
A Git worktree is a second working folder attached to the same repository. Git's own manual describes the feature as "manage multiple working trees attached to the same repository," which lets you "check out more than one branch at a time." git-worktree
That is the mechanism nearly every current ADE uses. Superset describes running "AI coding agents in parallel, each in its own isolated workspace," with a worktree per task. Superset documentation Orca's walkthrough gives each agent in a session its own worktree with its own branch checked out. Orca first session
A worktree is not a sandbox, and this is the most common misunderstanding of the model. Dure's documentation states that worktrees have "separate working directories and branches" but that they "can still produce conflicting changes, and they share the machine's credentials, processes, ports and network access." Parallel work · Current limits
Worktrees also inherit Git's own rules. A branch can be checked out in one place at a time, which is why OpenAI's worktree page documents the error you get when a second task wants a branch another worktree already holds. Codex worktrees
Sessions: work that outlives the window
A session is the running agent process plus its conversation. Because it is a process, closing a window and losing the work are different events, and an ADE is the component that has to tell them apart.
The honest limits are documented. Anthropic's background sessions "are preserved across sleep but stop if the machine shuts down." Agent view Dure's recovery "depends on saved state and provider support; it does not preserve a process across reboot," and its documentation adds the inverse warning: "a disconnected app is not evidence that its agent stopped." Current limits
Read that second sentence before you re-run a command. The failure mode of a session model is not usually lost work; it is the same task running twice.
Review: the diff and the reply
A diff is the list of lines a change added, removed or altered. Reviewing an agent's diff is a different activity from reading its transcript: the transcript says what the agent believes it did, and the diff says what is actually in your files.
Dure's review loop is to "read local diffs, send specific comments, and verify the result before combining work." Review and feedback Its built-in diff covers local worktrees, and changes made over SSH go through remote Git instead. Current limits
Attention: knowing which agent is waiting
With one agent, you are looking at it. With six, the scarce resource is your attention, and the workspace has to spend it for you. Anthropic's agent view sorts sessions by what they need rather than when they started, putting "Ready for review" and "Needs input" above "Working" and "Completed." Agent view
How is an ADE different from an IDE or a terminal multiplexer?
| IDE | Terminal multiplexer | Agent CLI | ADE | |
|---|---|---|---|---|
| What you open first | A file or project | A shell | A task, in one directory | A task, with its own checkout |
| Who writes the code | You, with suggestions | You | The agent | The agent |
| What it keeps separate | Nothing, by design | Panes and windows | Nothing, by default | A worktree and branch per task |
| Where review happens | In the editor | Wherever you run git diff |
In the transcript, or a /review command |
In a diff beside the agent |
| What it survives | Your own edits | A detached client | Its own process | A closed window |
The rows matter more than the columns, because real setups mix them. An ADE that cannot open your editor is not useful, and an editor is where most people still want to read code.
When do you not need one?
Three honest cases, in order of how often they apply.
You run one task at a time. Then the agent CLI alone is the whole answer, and a second layer only adds a thing to learn.
Your agent already includes one. This is the case most roundups skip. Anthropic's desktop app lets you "review diffs visually, run multiple sessions side by side, schedule recurring tasks, and start cloud sessions," and its background sessions move into an isolated worktree under .claude/worktrees/ before editing files. Claude Code overview · Agent view Codex offers worktrees in the ChatGPT desktop app, with setup scripts and a .worktreeinclude file for ignored files. Codex worktrees If you use one provider on one machine, check what you already have before adding anything.
You are already comfortable in a terminal multiplexer. If you can find the task you want without reconstructing where it lives, a visual workspace is a preference rather than a fix.
A separate ADE starts to earn its place at the point where those three stop holding: more than one agent CLI, more than one machine, or a layout that has to outlive a single task. Our companion guide works through that choice tool by tool: IDE, CLI, or ADE?
How can you test whether you need one?
Do this on a repository you know well, in one sitting.
- Start two small, genuinely independent tasks with the agent you already use, each in its own worktree.
- Walk away for ten minutes. Come back and time how long it takes you to answer one question: which task needs me first?
- Open the diff for the one that finished. Send one specific comment about one line, then confirm the agent received it.
- Merge both branches and run your checks on the combined result.
Step 2 is the measurement. If finding the waiting task was instant, you do not have the problem an ADE solves. If you reconstructed it from terminal tabs and your own memory, that is the cost a workspace is for. Keep the repository, agent and permissions identical if you later compare two tools, and treat this as your own evaluation rather than a benchmark.
Questions readers ask
Is an ADE the same thing as an agentic IDE?
Not as the sources use the terms. Warp's distinction is about who acts: an AI-assisted editor gives "inline suggestions and chat while a human writes most of the code," while an agentic environment lets an agent run multi-step tasks under supervision. Warp A forked editor with agent features can be either, depending on how you work in it.
Does an ADE replace my editor?
No. Isolation, sessions and review are the gap it fills; reading and writing code by hand is still an editor's job. Most of these tools open a worktree in whatever editor you already use.
Does an ADE come with its own model or subscription?
Not in this category. Dure's documented requirement is an "installed CLI and provider account," and it notes that "features vary by provider/runtime." Current limits Your agent subscription and its usage charges stay where they are.
Does running more agents finish the work faster?
Not automatically, and two documented costs push the other way. Anthropic notes that background sessions consume subscription usage like interactive ones, so "running ten agents in parallel uses quota roughly ten times as fast as running one." Agent view Dure's guidance is to add parallel work only "when the tasks can proceed independently," and to narrow scopes when two tasks start touching the same files. Parallel work Review is also serial: you still read every diff.
A new ADE appears every week. What actually separates them?
Not the worktree model, because they all use it. Four questions do the separating: how many different agent CLIs run in one window at the same time; whether an account belongs to the app or to the individual agent; whether one layout can span several projects and remote hosts, or one task fills the screen; and what happens after a disconnect or a reboot. The tool-by-tool guide applies those four to the current options.
Can an ADE keep an agent running after my machine shuts down?
Not on its own. Local sessions stop with the machine in both Anthropic's documentation and Dure's, and a reboot ends the original process. Agent view · Current limits Work that has to continue while your laptop is closed belongs on a cloud or remote runtime, which is a separate decision from the workspace.