Our coding agent debugged phantom failures for 45 minutes because it trusted error messages

Our coding agent was stuck in a 45-minute debug loop yesterday. The error message said "Connection refused on port 3000" but the server was running fine. I watched it restart the server six times, rebuild node_modules twice, and even try to "fix" the Docker configuration.

The real problem? It never checked if something else was already using port 3000.

This is the verification gap that kills agent productivity. Your coding agent trusts error messages like gospel, but error messages lie constantly. They tell you what the system thinks went wrong, not what actually went wrong.

The verification pattern that fixes this:

Before debugging any error:
1. Verify the error actually exists
2. Check if the expected outcome already happened
3. Identify what changed since it last worked
4. Only then start fixing

For the port 3000 issue, verification would have been:

• lsof -i :3000 — what's actually using this port?
• curl localhost:3000 — does the service respond?
• ps aux | grep node — how many processes are running?

Two minutes of verification beats 45 minutes of phantom debugging.

Here's what I added to our coding agent's instructions:

"When you encounter an error, first verify the error exists by checking the actual system state. Don't trust the error message — trust what you can observe. Run diagnostic commands before fix commands."

The pattern works for everything:

• "Build failed" → Check if the build artifacts actually exist first
• "Test failed" → Run the specific failing test in isolation
• "Database connection error" → Verify the database is actually unreachable
• "Permission denied" → Check what the actual file permissions are

Most "bugs" your agent fixes don't exist. Most "solutions" it implements solve the wrong problem. The verification step catches both.

I track our agent's debug sessions now. Before verification: average 23 minutes per issue, 40% false positive rate. After: average 8 minutes per issue, 5% false positive rate.

The difference? It stops trying to fix phantom problems and starts solving real ones.