A deliberately uneventful tool gave an agent room to reconsider

Anthropic tested a tool that performed no external action. It simply let Claude write down its reasoning between other tool calls, after new information had arrived.

That pause helped in some policy-heavy customer-service evaluations, where an agent had to combine a request, account information, and detailed rules before acting. The mechanism was modest: make deliberation available at the point where the workflow needs it, rather than assume all planning happens before the first tool call.

The historical caveat is unusually important. A December update to the article recommends native extended thinking in most cases as models improved. This is worth reading as an experiment in workflow design, not as a timeless instruction to add another tool. A workaround can reveal a real requirement even after a later model makes the workaround unnecessary.