The Workshop · 11 min mission
IDE Agents: The In-Editor Landscape
Map the in-editor coding agents — what agent mode is, who leads in 2026, and how they differ from CLIs.
On this page
An IDE agent is an autonomous coding loop inside your editor: you state an outcome in a chat panel, and the agent reads files, edits across the codebase, runs terminal commands and tests, reads errors, and iterates until done — rendering inline diffs you review. This guide maps the leading in-editor agents as of 2026-06-15, the exact commands/flags/prices to operate each, and how the IDE family relates to terminal CLIs (Claude Code, Codex CLI, Antigravity CLI).
The three interaction tiers
Every in-editor assistant exposes the same three tiers, named differently per vendor. The tier sets what the assistant may touch. Completion / Tab is inline autocomplete accepted with Tab (read-only awareness). Ask / Chat is single-turn Q&A about code (read-only by default). Agent mode is the autonomous tier: the model is handed tools (read_file, edit_file, run_in_terminal, search, MCP) and loops until the task completes. GitHub describes agent mode as an "orchestrator of tools like read_file, edit_file, and run_in_terminal" that runs commands and tests and auto-corrects in a loop.
| Tier | What it can touch | Vendor names |
|---|---|---|
| Completion / Tab | Suggests next keystrokes; accept with Tab. Read-only. | Copilot completions, Cursor Tab, Devin Desktop Tab |
| Ask / Chat | Single-turn Q&A; read-only by default. | Copilot Chat (Ask), Cursor Ask, Cline Plan |
| Agent mode | read_file, edit_file, run_in_terminal, search, MCP. Loops until done. | Copilot Agent, Cursor Agent, Cline Act |
The leading IDE agents, 2026-06-15
Cursor — standalone agent-first editor since the 3.0 rewrite (2026-04-02), no longer a VS Code extension. Current build v3.7 (changelog 2026-06-05). Open the parallel-agent surface with Cmd+Shift+P → Agents Window; it runs agents locally, in worktrees, in the cloud, and over remote SSH. In-house model Composer 2.5 (2026-05-18); also offers Claude, GPT-5.x, and Gemini 3.x. Rules live in .cursor/rules/*.mdc and a root AGENTS.md.
GitHub Copilot — agent mode is GA in VS Code, JetBrains, Eclipse, and Xcode. Two distinct surfaces: agent mode (synchronous, in-IDE) and the Copilot coding agent (asynchronous, cloud — see the callout below). Repo instructions read from .github/copilot-instructions.md, path-specific .github/instructions/NAME.instructions.md, and AGENTS.md (nearest in the tree wins).
Devin Desktop — the rebranded Windsurf (Cognition signed the acquisition 2025-07-14; rebrand shipped 2026-06-02). Default local agent Devin Local is a from-scratch Rust rewrite (up to 30% more token-efficient, with sandboxing and subagents) that replaces Cascade (EOL 2026-07-01). In-house model line SWE-1.6.
Cline — open-source (Apache-2.0, ~63.3k GitHub stars) across VS Code, JetBrains, a CLI (v3.0.24, 2026-06-11), and a web Kanban board. Bring-your-own-API-key: you pay Anthropic, OpenAI, Google, Bedrock, Azure, or local Ollama directly. Rules live in .clinerules at project root.
| Tool | Surface | In-house model | Cost model | Status |
|---|---|---|---|---|
| Cursor | Standalone editor (since 3.0) | Composer 2.5 | Subscription (Pro $20/mo) | v3.7, active |
| GitHub Copilot | VS Code / JetBrains / Eclipse / Xcode | MAI-Code-1-Flash | Subscription + AI Credits | Agent mode GA |
| Devin Desktop | Standalone editor (ex-Windsurf) | SWE-1.6 / Devin Local | Subscription | Rebranded 2026-06-02 |
| Cline | VS Code / JetBrains / CLI | None (BYO key) | Pay your provider directly | CLI v3.0.24, active |
| Roo Code | VS Code extension | None (BYO key) | Was BYO key | Archived 2026-05-15 |
| Tool | Price tiers | Instruction file(s) |
|---|---|---|
| Cursor | Hobby free · Pro $20/mo · Pro+ $60/mo · Ultra $200/mo · Teams $40/user/mo | .cursor/rules/*.mdc, AGENTS.md |
| GitHub Copilot | Free $0 · Pro $10/mo · Pro+ $39/mo · Business $19/user/mo · Enterprise $39/user/mo | .github/copilot-instructions.md, .github/instructions/*.instructions.md, AGENTS.md, CLAUDE.md/GEMINI.md |
| Cline | No subscription — pay model provider directly (optional hosted Cline credits) | .clinerules |
| Google Antigravity | Free preview (rate-limited Gemini) · AI Ultra $100/mo (5× Pro) · top tier $200/mo (20× Pro) | AGENTS.md |
Google Antigravity
Google Antigravity launched 2025-11-18 alongside Gemini 3 as a free, agent-first IDE built as a heavily modified fork of VS Code (macOS 12+, Windows 10+, 64-bit Linux). It has two surfaces: an Editor view (traditional IDE plus an agent sidebar) and a Manager surface for spawning and orchestrating many agents asynchronously across workspaces. Agents produce verifiable Artifacts — task lists, implementation plans, screenshots, browser recordings — and have direct access to the editor, terminal, and an integrated browser.
Antigravity 2.0 (Google I/O, 2026-05-19) added multi-agent orchestration, scheduled background tasks, an SDK, native voice, and a new CLI built in Go that is absorbing Gemini CLI. The default coding model became Gemini 3.5 Flash. The public preview stays free for individuals on rate-limited Gemini access.
IDE agents vs terminal CLIs
IDE agents (Cursor, Copilot agent mode, Devin Desktop, Cline, Antigravity) and terminal CLI agents (Claude Code, Codex CLI, Antigravity CLI) run the same agentic loop — plan, edit, run, read errors, iterate, with tool access. They differ in surface: the IDE renders the loop visually (inline diffs, file tree, gutter marks, integrated debugger/terminal, a built-in browser for UI checks); the CLI runs in your shell and prints diffs as text you can pipe, script, and run headless in CI.
Where each surface is strongest
IDE agent (in-editor)
Watch the loop happen. Inline diffs accepted hunk-by-hunk, a live file tree and debugger, a built-in browser for UI verification, and Tab completion as you type.
Best for: tight, reviewed changes where you stay in the file and want to see each edit land.
Terminal CLI agent
Hand off a goal. Runs headless in CI, on a schedule, or fanned out across many parallel sessions; prints diffs as text you can pipe and script.
Best for: autonomy and scale — long jobs, automation, and background work returning a diff or a PR.
The convergence: three shared layers
The IDE-vs-CLI split is becoming a UI preference, driven by three concrete shared layers.
Shared instruction file. AGENTS.md is read by Cursor, GitHub Copilot, and the CLIs alike (Copilot also reads CLAUDE.md and GEMINI.md). Write build commands and constraints once; most agents pick them up.
Shared models. The same frontier models — Claude Opus 4.x, GPT-5.x, Gemini 3.x — back IDE agents and CLIs. The vendor's in-house model (Cursor Composer, Devin SWE-1.x) is the differentiator, not the headline engine.
Shared protocol. The Agent Client Protocol (ACP) — an open, Apache-2.0 agent↔editor standard — lets a CLI agent plug into an editor as a first-class citizen. Devin Desktop adopts it via an Agent Command Center, running Codex (CLI), Claude Agent, OpenCode, Junie, and Gemini CLI inside the editor alongside Devin's own agents.
Knowledge check
A teammate says "just install Windsurf with Cascade" and "grab Roo Code, it’s great." It is 2026-06-15. What is the accurate correction?
Which IDE agent fits you?
Which IDE agent?
Five quick questions about where you code, how much autonomy you want, how you'd rather pay, whether open source matters, and your main language — and you'll get a single agent to start with, the reasoning behind it, and a runner-up. All five contenders are in the legend below.
The five contenders
An AI-first editor (a VS Code fork) with a strong agent, fast Tab completion, and multi-file edits.
Drops into the editor you already use — VS Code, JetBrains, Visual Studio, Neovim — with chat, agent mode, and completions.
An AI-native editor whose "Cascade" agent keeps the whole task in flow — opinionated, polished, and beginner-friendly.
An open-source, fully autonomous coding agent that lives in VS Code. You bring your own model and pay the provider directly.
Open-source VS Code agent with role-based "modes" and deep auto-approve control — the tinkerer's pick. BYO model.
Where to go next
Deeper guides in this constellation cover each thread with the exact commands and config: Cursor essentials and Cursor rules & context; GitHub Copilot agent mode for the synchronous-vs-cloud split in depth; Windsurf / Devin Desktop for the rebrand and Devin Local; Cline & Roo Code for the BYO-key path and Roo's mode architecture; IDE agents vs CLI agents to pick a surface per task; and rules files across tools to make AGENTS.md work everywhere at once.
Reach the end and this star joins your charted sky.