The Autonomous Dev Tool Wars

Yesterday afternoon, Anthropic shipped Opus 4.7 with a 1 million token context window in beta. That’s the biggest number any production coding agent has hit. It landed while I was mid-test on a client refactor, and I was the person who had to decide, live, whether to switch tools in the middle of the job.

Two weeks earlier, OpenAI dropped the ChatGPT 5.5 desktop super app with Codex, the Atlas browser, and GPT-5.4 chat in one window. Three weeks before that, Cursor turned on 8 parallel background agents. Back in December 2025, Cognition (the company behind Devin) bought Windsurf for around $250 million.

Five major autonomous coding agents. 90 days. That tempo is not normal. Something has shifted in how code gets written, and most people are still typing one line at a time.

This issue is the map I wish I had in January. What the shift actually is, what the five tools each do, how to chain them, and a 90-day plan to get on the right side of it.

The shift from pair programmer to teammate

Anthropic, GitHub, and Cursor all quietly renamed their products in early 2026. The old framing was “AI pair programmer”. You typed. It suggested. You accepted or rejected.

The new framing, the one all three companies now use in their own marketing, is “AI teammate”. You delegate. It runs. It comes back with a pull request.

That sounds like a small change. It isn’t. Pair programming keeps you in the driver’s seat for every keystroke. Delegation means you’re reviewing work from an agent you dispatched an hour ago. Different skill. Different workflow. Different tools.

The data backs the shift. GitHub’s own numbers say more than 51% of code committed to GitHub in early 2026 was AI-generated or AI-assisted. In early 2024, that number was close to zero. In 24 months, it went from experiment to majority. The AI coding tools market is around $12.8 billion in 2026, up from $5.1 billion in 2024.

The job is no longer writing code alongside AI. It’s reviewing pull requests from agents you dispatched.

The system: five tools, one stack

Here are the five I’ve been running against real client work for the last two weeks. I’ll spend one paragraph on each, and then show you how to chain them.

Claude Code 4.7. Anthropic’s terminal-native coding agent, paired with Opus 4.7 (1M context, in beta). You install with one npm command. It runs in your terminal, talks to MCP servers (the protocol that lets the agent use external tools like Pinecone for search or Playwright for browser control), runs subagents with isolated contexts, and supports hooks that run shell commands on specific events. I use it as the primary orchestrator. $20/month Pro, $100/month Max 5x, $200/month Max 20x.

OpenAI Codex (the 2026 version). Cloud-based software engineering agent bundled into ChatGPT 5.5. Different from 2021 Codex. You give it a task, it runs in a cloud sandbox, parallelizes across multiple jobs, returns pull requests. Included with ChatGPT Pro, Business, Enterprise.

Perplexity Comet and Computer. Comet is the AI browser (free, $20/month Pro). Computer is a desktop autonomous agent at $200/month that treats the browser as its primary tool. I use Comet for non-code research: comparing 12 SaaS pricing pages, filling vendor forms, pulling from 20 tabs into one summary. Computer is for long-running multi-tab jobs.

GitHub Copilot Workspace. Evolved from the autocomplete tool of 2022. You assign a GitHub issue to Copilot. It reads the issue, writes the code, runs tests, opens a pull request. Free tier available; Pro $10/month, Business $19/seat/month. If your team lives in GitHub issues, this is the most mature issue-to-PR agent available.

Cursor. VS Code fork with three modes: Composer for multi-file coordinated edits, Agent for autonomous task execution in the IDE, and Background Agents for up to 8 parallel cloud agents. Pro $20/month. If you live in the editor and want the visual diff review, Cursor beats everything else.

Bonus: Windsurf. Codeium’s AI-native IDE, acquired by Cognition in December 2025. $15/month Pro, cheapest in the set.

Implementation: the 90-day plan, compressed

The full 90-day roadmap is in the playbook (link at the bottom). Here’s the short version.

Weeks 1 to 4: master Claude Code first. Resist the urge to install all five tools in week one. The compounding benefit of one tool learned deeply beats four tools used shallowly, every time. Set up your claude.md file (mine is 400 lines: voice rules, banned vocabulary, project conventions). Build your first subagent. Write your first skill. Connect one MCP server.

Weeks 5 to 8: add Cursor for IDE work. Claude Code gets anything over 20 minutes or involving browser automation. Cursor gets anything under 20 minutes where you want the visual diff.

Weeks 9 to 12: add a third tool based on your job. Solo builders add Comet. Agency teams add Copilot Workspace. Researchers add Perplexity Computer.

By day 90, the target is a measurable 2x improvement in time-to-ship on comparable work. That’s what happened on my own work between January and April.

Three stack configurations

• Solo builder stack. Claude Code Pro + Cursor Pro + Comet free. About $40/month. Ships at $50/hr rate, breaks even in 3 hours a week saved.

• Agency team stack. Claude Code Enterprise + Copilot Business + Windsurf Pro. About $135/seat/month. Break-even at 2 hours per developer per week.

• Researcher stack. Claude Code Max + Perplexity Computer + Codex via ChatGPT Pro. About $500/month. Only worth it for heavy research-and-writing roles.

Three pitfalls I hit and want you to skip

Pitfall 1: picking a tool per project instead of per task. People marry one tool and force every job through it. Wrong. Different tools win different tasks. Pick one per task.

Pitfall 2: skipping Plan Mode. The biggest time sink in agent work is when the agent does the wrong thing confidently. You come back an hour later and find it built the feature in the wrong layer. Plan Mode makes the disagreement visible before any code runs. I use it for anything touching more than 3 files.

Pitfall 3: running all 5 tools simultaneously. I tested this for a week and lost around 2 hours just switching between UIs. Three tools is the sweet spot. Four is the max before diminishing returns.

I’m writing this one day after Opus 4.7 shipped, so some numbers here will shift over the next few weeks. The frame won’t. The shift is real, the tools are here, and the people who learn to chain them will ship what two-person teams shipped last year.

PS. I put together a 40-page playbook covering everything I learned: the full decision tree, the cost math, the three stack configurations, and the 90-day roadmap. Reply to this with “STACK” and I’ll send you the link.