AI Stack Layer · 5 of 8

Dev Agents & AI IDEs

AI that reads, writes, and edits real codebases — from autocomplete that finishes your line to autonomous agents that ship features. The fastest-changing layer of the stack.

Code GenerationPair ProgrammingAgentic CodingIDE-nativeLayer 5
← Back to AI Landscape
Quick Facts

At a Glance

Basic Concepts

  • Three modes: autocomplete (inline), chat (sidebar), and agent (autonomous multi-file edits).
  • Context is everything — the agent needs to see the right files to make good edits.
  • Tool use: read files, run shell commands, search the web, call MCP servers.
  • Review loop: AI proposes → you approve → tests run → iterate.
  • The model matters: the same IDE feels totally different on Claude vs GPT vs a local model.
Landscape

The Major Tools

ToolForm factorWhy it stands out
Claude CodeTerminal / IDE pluginAnthropic's official agent — runs in your shell, edits files, runs tests, full agent loop.
CursorVS Code forkMost popular AI-native editor; Composer agent + Tab completion.
GitHub CopilotPlugin (any IDE)The original. Inline + chat + agent (Workspace).
Windsurf (Codeium)VS Code forkCascade agent; strong on multi-file flow.
Cline / Roo CodeVS Code extensionOpen-source agentic coding inside VS Code.
AiderTerminal CLIOSS agent that edits via git commits.
Zed AINative editorRust-built collaborative editor with AI baked in.
JetBrains AI AssistantJetBrains pluginNative to IntelliJ / PyCharm / WebStorm / Rider.
Devin / Replit Agent / Bolt / LovableCloud workspace"Type a prompt, get an app" — async cloud-based agents.
Tabnine / Supermaven / CodeiumInline completionFast, often local-first, cheap.
Mechanics

How They Work

The Three Modes
  1. Autocomplete: ghost-text suggestions as you type. Latency < 200ms; small or distilled models.
  2. Chat: sidebar conversation, with the open file (or workspace) as context.
  3. Agent / Composer: the AI plans, edits multiple files, runs commands, iterates — you approve.
Context Strategy

Modern coding agents work surprisingly well because they aggressively gather context:

  • Repo indexing (embeddings of every file).
  • Static analysis (LSP — definitions, references, types).
  • Recent edits, open tabs, terminal output.
  • Web search for libraries / errors.
  • MCP servers for external tools (Jira, DBs, Sentry).
The Model Context Protocol (MCP)

MCP (Anthropic, 2024) is an open standard for connecting AI agents to external tools — file systems, databases, ticketing, monitoring. Pluggable across Claude Code, Cursor, Windsurf, and many others.

What AI Coding Is Good / Bad At
Good at
  • Boilerplate, tests, docstrings.
  • Translating between languages / frameworks.
  • Refactors with clear before/after.
  • Debugging stack traces.
  • "Do the thing in 3 places" repetitive edits.
Bad at (still)
  • Architecture decisions across many files without context.
  • Domain-specific business rules nowhere documented.
  • Performance optimization without measurement.
  • Recognizing it doesn't know.
Picking a Tool
If you want…Try
Best agentic coding todayClaude Code, Cursor (Composer), Windsurf
Stay in VS Code, official-feelingGitHub Copilot
Stay in JetBrainsJetBrains AI Assistant, Junie
OSS / self-hostAider, Cline, Continue.dev
Async cloud agentsDevin, Replit Agent, Bolt, Lovable
Just fast autocompleteSupermaven, Codeium, Tabnine
Continue

Other AI Stack Layers

Foundation ModelsModel ProvidersFrameworksVector DBsMLOpsClassic MLData Prep ↑ Back to AI Landscape