Since I moved to my own Agent, I’ve seen higher token consumption for the same tasks than what Claude Code uses. This was in my head for a while, so I decided to investigate.
First, I added a took that tracks all iteractions in a sqlite3 database, which with a few days of data it was possible to see that my scout subagent was consuming most of the tokens.
Then I asked it to compare with the explore agent from the leaked claude code source, and found the issue: Claude was using Haiku without thinking, and I was using Sonnet with high thinking.
Fun debug session, now the token consuption is controlled.