Agent Development Environment

An agent development environment, or ADE, is the place where an AI agent gets built, run or tested. The catch is that people use the phrase for three different things.

Example

A researcher wants to know whether an agent can fix broken code. She sets up a task: a clean copy of the code, plus hidden checks that pass only if the fix works. The agent edits files inside a sandbox, a walled-off computer where its changes can't touch anything else. She scores the fix, resets everything, and runs it again.

That one scene has three layers. The task and its hidden checks are the test environment. The sandbox is the runtime. The code that tells the agent what to do next is its harness.

The three meanings

• A coding-agent product. An app and set of tools that help a developer change software. Claude Code is one.
• A training or test environment. The tasks, what the agent can see, what it's allowed to do, and how it gets scored. Reinforcement learning, which trains a model by scoring its attempts, runs on these. Verifiers is a toolkit for building them.
• An execution runtime. The computer, and the walls around it, where an agent's tools actually run. A hosted sandbox like the ones in Modal's documentation handles where the code runs, not whether the task is any good.

An agent framework is a separate layer again. It coordinates the model calls, the tools and what the agent remembers between steps. It can use any of the three above without replacing them.

To choose between these layers, read the full ADE guide. For real projects, browse the RL environments directory. For the bigger idea, see Agentic AI.