AI-Powered IDEs, Harnesses, and Your Dev Environment

A year ago, this chapter would have been called “Cursor.” You’d install it, set up .cursorrules, and your AI coding environment was done.

That’s not the world we live in anymore.

In April 2026, the developers I respect most are running five to fifteen AI sessions in parallel. Boris Cherny, the creator of Claude Code at Anthropic, runs 10-15 concurrent agents, shipping 20-30 pull requests per day. Mitchell Hashimoto, the founder of HashiCorp, maintains four parallel checkouts and gives the same task to different agents as a competition. A JetBrains survey of over 10,000 developers found that 85% use AI coding tools regularly and 70% stack two or more tools simultaneously. The tooling revenue hit $12.8 billion in 2026.

Two ways to work with AI on code

Understand the fundamental split before looking at specific tools.

Collaboration-first is what Cursor and Copilot do. You’re in the editor. You’re driving. The AI suggests completions, answers questions, makes inline edits. You approve every change.

Delegation-first is what Claude Code and Devin do. You describe a task. The agent plans, edits files across your codebase, runs commands, tests its work, and iterates until it’s done. You review the result, not the process.

Neither is better. They’re different tools for different shapes of work. The best setup in 2026 uses both.

One finding that surprised me: the same AI model scores dramatically differently depending on which tool wraps it. Claude Opus 4.5 scored 77% in one harness and 93% in another on the same benchmark. I spent months thinking the model was the variable that mattered. It’s not. The wrapper, the context engineering, the permission system, the verification loop — those determine whether you get good code or garbage.

Cursor

Cursor is an AI-native code editor from Anysphere, a fork of VS Code. Over a million daily active users. Two billion dollars ARR by February 2026. 67% of the Fortune 500.

Tab completion is the strongest feature — roughly 72% acceptance rate versus Copilot’s 35-40%. Cmd-K inline edits let you describe changes in plain English and get a diff. Agent mode edits across multiple files and runs terminal commands. Background Agents work in cloud VMs, opening PRs while you do something else. Design Mode lets you annotate UI elements directly in a browser and the agent implements changes.

The pricing mess: Cursor Pro is $20/month but credit-based since June 2025. Heavy users hit their limit mid-month. Anysphere apologized and issued refunds. The trust damage was real. Two of our engineers switched to Cline after hitting the credit ceiling three months in a row. They’re happier and spending less.

VS Code + GitHub Copilot

4.7 million paid subscribers. 90% adoption in the Fortune 100. Agent Mode went GA in March 2026. The Coding Agent takes a GitHub issue and returns a finished PR. Multi-model support lets you switch between Claude, GPT-5, and Gemini per session.

Free tier gets 50 premium requests and 2,000 completions per month. Pro at $10/month. Important: in April 2026, GitHub paused new sign-ups for individual Pro plans, citing inference cost pressures. A significant signal about the economics of unlimited AI access at $10/month.

When Copilot beats Cursor: cost, ecosystem integration, IDE breadth. When Cursor beats Copilot: tab completion quality, codebase context depth, more cohesive AI-first experience.

The open-source agent family

The biggest gap in most guides. Cline, Roo Code, and Kilo Code are free, open-source, bring-your-own-key VS Code extensions that turn your editor into an agentic environment.

Cline (5 million installs, 58K GitHub stars) pioneered the Plan/Act split — read-only analysis before making changes. Its MCP marketplace is considered the best in open source. The trade-off is cost: bring your own API key, and heavy Claude usage runs $50-200/month.

Roo Code is the most popular fork, differentiated by multiple modes (Architect, Code, Ask, Debug, Orchestrator) and subscription-bridge providers that route through your existing Claude Pro subscription to cut costs.

Kilo Code is the newest and most ambitious — $8 million raised, inline autocomplete (most Cline-family tools lack this), and JetBrains support alongside VS Code.

When to use these over Cursor: when you want full control of API costs, need a specific model (DeepSeek, local Ollama), or want the transparency of open source. The practical cost-conscious stack: Copilot for autocomplete ($10/month) plus Cline with your own API key for agentic work.

Terminal harnesses

Beyond Claude Code (18% workplace adoption, highest satisfaction scores), the field includes Codex CLI (open-source, Docker-sandboxed, free for ChatGPT subscribers), Gemini CLI (open-source, 1M-token context, 1,000 free requests per day with just a Google account), Aider (43K stars, git-first with auto-commits, maximum control), and OpenCode (95K+ stars, 75+ model providers, the recommendation for cost-sensitive developers).

Vibe coding: for people who don’t write code

A quarter of Y Combinator’s Winter 2025 batch had codebases 95% AI-generated. Lovable reportedly hit $300M ARR by January 2026. Bolt.new goes from prompt to running app in under a minute. v0 is the best React component generator. Replit Agent is the most complete platform.

The honest assessment: these are real for landing pages, internal tools, and MVPs. But AI-generated code has 2.74x more XSS vulnerabilities per CodeRabbit’s research, and architectural drift compounds as projects grow. I’ve had three potential clients come to us in the last quarter with Lovable-generated apps that worked beautifully for the demo and fell apart under real traffic. The code wasn’t bad. It just wasn’t built for what came next. Use them for the first version. Bring in proper engineering for what follows.

AGENTS.md and the infrastructure layer

Every major coding agent now reads a project-level instruction file. The Linux Foundation placed AGENTS.md under the Agentic AI Foundation in December 2025. Over 60,000 repositories ship one.

Keep it under 200-300 lines — research showed longer files actively degrade agent performance. Include your tech stack, build commands, hard rules, code style, and most importantly: mistakes the agent has made before. That last one is the real power. Cherny’s team updates theirs multiple times a week. Every time an agent does something wrong, they add a rule. The instruction file is collective memory. It gets better every day.

MCP servers connect agents to external services (covered in Chapter 12). The essential ones: GitHub MCP for issues and PRs, Context7 for real-time library documentation (fixes “the AI hallucinated a deprecated hook”), and Playwright MCP for verifying UI changes locally.

The cautionary data

Anthropic’s own research found developers using AI scored 17% lower on code comprehension quizzes than those who coded by hand. Multiple open-source projects cracked down on AI-generated PRs in 2026. cURL shut down its bug bounty program after six years because of AI-generated noise.

The tools make you faster. They might also make you shallower. The antidote is the discipline that’s always mattered: understand what you’re shipping. I make our junior engineers explain their AI-generated PRs line by line during review. If they can’t explain why a particular approach was chosen, the PR goes back. The tool wrote the code. They need to own it.

The bottom line

Your AI development environment in 2026 is not one tool. It’s a portfolio. An IDE for visual editing. A terminal harness for autonomous work. An instruction file that teaches every agent your project’s rules. MCP servers that connect agents to your infrastructure.

The wrappers will keep changing. Cursor might be owned by xAI by the time you read this. GitHub paused new Copilot sign-ups the same week. New tools appear monthly. What doesn’t change: the discipline of building a good harness around your agents. Write a tight instruction file. Update it every time something goes wrong. Verify output with real tests. Run sessions in parallel. Understand what you’re shipping, even when you didn’t write every line.

The workflow matters more than the model. The discipline is the durable skill.

---

This is the free web edition of Chapter 7. The full text — with IDE comparison matrices, terminal harness benchmarks, AGENTS.md templates, MCP server configurations, and practitioner workflow breakdowns — is available in 42: The AI Builder’s Stack, coming Q3 2026 on Amazon in hardcover, paperback, and digital.