Why an IDE Is Not Enough: Building an ADE (Agentic Development Environment)

An IDE organizes an individual's work on code, while an Agentic Development Environment (ADE) must also manage the work of autonomous agents, which can execute in parallel, utilize different models, and build context over time. When developing an ADE, the focus shifts from integrating AI chat within an editor to turning agent execution into a structured, planable, and reviewable form of engineering work.

An IDE operates within a single session, with much of its state residing in the local session, including open files, terminal, current branch, and editor history. Closing a terminal terminates the operational context, while opening a new machine creates a new session. In contrast, an ADE requires work to persist beyond the session.

A card in an ADE serves as the unit that connects intent, execution, and evidence, becoming more than a visual Kanban item. It includes the goal, acceptance criteria, context and decisions, selected agent and model, execution, logs, artifacts, result, and optional review.

The difference between an ADE and an IDE lies in the interface and how users interact with the system. Users should not need to remember which terminal initiated an investigation, nor should they have to understand what an agent did or what failed. Instead, they should be able to inspect a card and comprehend the request, the agent's actions, any failures, and the next steps.

In an ADE, work is split into independent units that can run in parallel, but they must also be connected enough so that crucial decisions are not lost. In the AuraPunk ADE, cards and SPECs (Specification cards) provide work identity before execution begins. This approach enables the split of work into distinct units, each assigned to a specific agent, facilitating parallel execution.

For instance, one agent investigates the codebase, another proposes or implements a bounded change, another writes or validates tests, and another handles ambiguous decisions using a stronger model. The ADE ensures these executions remain connected to the same project, assigning agents based on their capabilities and suitability for the task at hand.

An ADE is designed to work with various AI agents such as Codex, Claude Code, OpenCode, Qwen Code, Gemini CLI, Command Code, and Antigravity CLI. The goal is not to hide their differences but to orchestrate them effectively. Each agent has a profile comprising reasoning quality, speed, cost, available tools, interaction style, and ability to operate within a codebase.

The agent choice should be made at the card level, based on the specific task requirements, ensuring the right agent and model are utilized for the appropriate type of work. This enables flexibility, cost-effectiveness, and efficiency in the development process.