We stopped using LangGraph. Here's what runs our agents now.

You're staring at three coding tools and wondering which one to use. Wrong question.

The right question is: what kind of coding problem are you solving?

After running all three for six months, here's the decision tree that actually matters:

Use Cursor when: You need tight feedback loops on existing code. Its agent loops are sharp — it sees your cursor position, understands file context, and can iterate on the same function 15 times without losing track. Perfect for refactoring, debugging, or polishing features.

Use Claude Code when: You're building something new or complex. Its context window and reasoning are unmatched for architecture decisions, multi-file changes, or when you need to explain your reasoning to someone else later. SWE-bench scores don't lie.

Use Codex when: You need speed over sophistication. Autocomplete, simple functions, boilerplate. It's fast but will miss edge cases and sometimes generate syntactically correct nonsense.

But here's what changed everything: we stopped picking one.

Our coding agent now routes automatically:

if task.involves_multiple_files() and task.complexity > 7:
    route_to_claude_code()
elif task.is_iteration() and cursor_context_available():
    route_to_cursor()
else:
    route_to_codex()

The magic happens in combination. Claude Code writes the architecture and complex logic. Cursor refines it with tight feedback loops. Codex fills in the boilerplate.

We track three metrics:

• First-pass success rate: Does it work without human intervention?
• Token efficiency: Cost per working line of code
• Iteration depth: How many back-and-forth rounds before completion?

Results after 200 coding sessions:

• Claude Code: 73% first-pass success, high token cost, 2.3 iterations average
• Cursor: 68% first-pass success, medium cost, 4.1 iterations (but faster iterations)
• Codex: 45% first-pass success, low cost, 1.2 iterations
• Hybrid routing: 81% first-pass success, 40% lower cost than Claude-only

The breakthrough was treating them as specialized tools, not competing platforms. Your coding agent should be a conductor, not a virtuoso on one instrument.

One more thing: quota limits force good routing discipline. When Cursor hits its limit, falling back to Claude Code direct API often gives you better quotas anyway. When Claude gets expensive, Codex handles the simple stuff.

Stop asking "which is best." Start asking "which is right for this specific task."