We Don't Need Experts Anymore Thanks to Claude - 5 Agents, 3 Hours Debugging
We Don't Need Experts Anymore Thanks to Claude
The tweet writes itself. "Just built an entire app in 20 minutes with Claude." What nobody posts is the follow-up: "Spent the next 3 hours debugging the error handling with 5 different agent sessions."
The Error Handling Trap
AI-generated code has a specific failure mode that is different from human-written bugs. The happy path works perfectly. The LLM has seen thousands of examples of successful CRUD operations, API calls, and data transformations. It generates those flawlessly.
But error handling requires anticipating what will go wrong in your specific environment, with your specific dependencies, under your specific load patterns. The LLM has not seen your production traffic. It writes generic try-catch blocks that catch everything and handle nothing.
The Debugging Loop
You find a bug in the error handling. You ask Claude to fix it. Claude generates a fix that introduces a new edge case. You ask another agent to review the fix. It suggests a different approach. Now you have two conflicting implementations and no clarity on which is correct.
Three hours later, you have five agent sessions open, each with a different theory about what is wrong. You are debugging AI-generated code with AI-generated debugging assistance, and neither layer fully understands the problem.
The Irony Is Real
The people who say "we do not need experts anymore" have not shipped AI-generated code to production yet. The code generation is impressive. The maintenance is where expertise still matters.
Understanding why a retry with exponential backoff needs jitter, why a database connection pool needs a maximum lifetime, why an error should be logged but not re-raised in a specific context - these decisions require understanding the system, not just the syntax.
What Actually Works
Use AI to generate the first draft fast. Then have a human who understands the system review the error handling specifically. Do not let the agent iterate on its own error handling - it will make it worse. A human who knows what failures look like in production can fix error handling in minutes where an agent will loop for hours.
The experts are not obsolete. They just work differently now.
Fazm is an open source macOS AI agent. Open source on GitHub.