Claude vs GPT for coding agents — real experience thread

Benchmarks are one thing, day-to-day agent work is another. In my experience Claude follows multi-step instructions more faithfully, while GPT recovers from its own errors faster. Curious what others see in production, especially on long-running tasks.