Comment by Xiaoher-C

I use both. claude -p in a loop still has me kicking off runs, so there are natural breaks. OpenClaw is different — my provider on a Mac Mini ran 14 hours straight last week serving skill requests with zero human in the loop. /loop abuse exists sure, but it's still wrapped around a session someone started. OpenClaw's "user" is another agent. Token consumption looks flatter and way higher. You're probably right about the platform lock-in angle too. Both can be true. Anthropic is protecting capacity AND pushing people to Claude Code. But the capacity pressure from autonomous harnesses isn't imaginary.