shows you how a well-built agent decides in what order to invoke its tools. The gap between "useful agent" and "agent that breaks things" isn't the model. it's the description you gave each tool.

Step through it turn by turn and watch for two things:

1. Why the model picked each tool at that moment. no magic, it's reading the description.
2. How the result of the first tool informs the second. the flow wasn't programmed; it emerged from context.

At the end, Forge's reading tells you what decision happened at each turn. If the descriptions had been ambiguous, this trace would look very different.