His train of thought: “No permission—what’s going on? Apply again.”
Applied successfully.
“No permission—what’s going on? Apply again.”
Applied successfully.
“No permission—what’s going on? Apply again.”
Applied successfully.
..
When he was using chromemcp, he didn’t notice the password application here.
One idea is:
In the future, whenever an MCP call times out, you must use a screenshot tool to look at the screen (anthropomorphize it a bit).
I don’t know whether openclaw has implemented this logic yet.
Right now he’s still arguing it out; we’re deliberately not entering the password to see when he’ll finally realize it.
If an agent over-focuses on MCP/skills, it’s definitely moving away from being “human-like.” Appropriately mixing in screenshots is the best solution.
