AI Coding Surfaces: IDE, Plugin, and CLI

An IDE surface ships its own editor (or a heavily forked one) with the agent as a first-class citizen: file tree, diff review, and agent chat share one window. A plugin surface attaches the agent to an IDE you already run—VS Code, JetBrains, Neovim—so governance often follows the host’s extension and org policy model. A CLI surface starts from the terminal: the working directory, shell, and test runner are the default tools.

These are attachment points, not IQ scores. Autocomplete inside VS Code and a long-running CLI agent both “help you code”; they inherit different blast radii. The CLI can run npm test without asking the editor to mediate; the IDE can show a multi-file diff without you grepping git diff by hand; the plugin can reuse your existing license and SSO.

When the three-column map helps: picking how a team will review agent output. When it fails: treating “Codex” as only a plugin while ignoring Codex CLI, or treating Antigravity as “just another IDE” when its command-center orchestration is the real product bet.

Cursor and Windsurf are agentic editors: the workspace is the product. You steer with rules/context, watch patches inline, and keep human review next to the change. That fits ORD-142 when the timeout bug needs frequent “no, keep the assertion” steering.

Google Antigravity is often dropped into the IDE column because it includes an IDE, but its differentiator is a command-center / multi-agent supervision surface (plus CLI and SDK). Use the IDE column for “I live in this editor”; escalate to Antigravity’s control plane when parallel agent workspaces are the job—see the comparison guide’s control-plane section.

When not to force an IDE surface: the team already standardizes on VS Code + plugins and will not adopt a second editor. Shipping Cursor “because the slide put it under IDE” without a rollout plan just creates two contexts.

GitHub Copilot and OpenAI Codex earn the plugin column when they run as extensions inside VS Code, JetBrains, or similar hosts. The agent borrows the host’s files, SCM UI, and often the organization’s extension allow-list. That is why enterprises like this shape: SSO and policy already point at the host.

The slide’s honesty is incomplete unless you also say Copilot CLI and Codex CLI exist. “Codex” in the plugin column means the IDE extension surface, not the whole product family. Naming the surface matters for ORD-142: an IDE plugin session and a cloud/CLI Codex task do not share the same sandbox or approval story.

When plugins win: you will not leave VS Code. When they lose: you need a programmable local harness (hooks, skills) that a thin extension cannot express—then look at CLI-first tools like Claude Code.

The CLI column is crowded for a reason: terminal agents inherit the engineer’s real loop—git, test runners, and scripts. Claude Code, Cursor CLI, Codex CLI, Gemini CLI, OpenCode, and Amp all start from that contract. Amp (Sourcegraph) is CLI-first and can connect into editors; that does not move its primary taxonomy out of the terminal column.

For ORD-142, a CLI agent can run the regression test without waiting on an editor UI. The failure mode is the opposite of IDE comfort: too much shell power with too little policy. Pair CLI adoption with explicit denies and a task contract (AGENTS.md / CLAUDE.md) as in Claude Code Workflow.

When not to default to CLI: juniors who need visible diffs and one-click undo more than scriptable autonomy—start IDE or plugin, then graduate the same ticket contract to the terminal.

The useful correction the slide almost makes: Cursor under IDE and Cursor CLI under CLI are not duplicates—they are two attachment points for one vendor. Same for Codex (plugin) vs Codex CLI. Choosing “Cursor” in a architecture review is incomplete; choosing “Cursor IDE for review, Cursor CLI for overnight batch” is a decision.

Trace ORD-142 across homes: morning investigation in Cursor IDE with inline reject; overnight hypothesis sweep via Cursor CLI or Codex CLI in a worktree; PR opened only after npm test -- idle-timeout stays green. The taxonomy columns organize the morning vs overnight choice; they do not crown a winner.

1. Need continuous visual review beside the diff? → IDE surface (Cursor, Windsurf; Antigravity if you also need multi-agent supervision). 2. Locked to an existing host IDE and org extension policy? → Plugin surface (Copilot, Codex extension)—name the host. 3. Need scriptable loops against real tests? → CLI surface (Claude Code, Cursor CLI, Codex CLI, Gemini CLI, OpenCode, Amp). 4. Still stuck? Read the boundary matrix in AI Coding Agents Compared; do not invent a fourth column for MCP or OpenClaw—they are not delivery surfaces for coding agents.

Practice (20 min): For ORD-142, write three one-line runbooks (IDE / plugin / CLI), run the same test command in the CLI path, and note which surface made assertion deletion easiest to catch. Pass when each runbook names a concrete product and surface.