Agentic tasks use up a huge amount of tokens compared to simple chatting. Every elementary interaction the model has with the outside world (even while doing something as simple as reading code from a large codebase) is a separate "chat" message and "response", and these add up very quickly.