Comment by baal80spam
Comment by baal80spam a day ago
> Use Claude 4.5 Opus
In my org we are experimenting with agentic flows, and we've noticed that model choice matters especially for autonomy.
GPT-5.2 performed much better for long-running tasks. It stayed focused, followed instructions, and completed work more reliably.
Opus 4.5 tended to stop earlier and take shortcuts to hand control back sooner.
a ralph loop can make claude go til the end, or to a rate limit at least.
opus closes the task and ralph opens it right back up again.
i imagine there's something to the harness for that, too