Before it
Tool use was taught with human demonstrations or elaborate prompting.
The problem
Demonstrations are expensive and do not transfer to new tools.
What it introduced
Let the model insert candidate API calls into text, keep the ones that reduce its own loss, and fine-tune on those.
Why it matters
It framed tool use as something a model can learn by itself from its own uncertainty.
What modern models inherit
Tool calls as tokens the model emits, rather than an external orchestration layer.
What came next
Function calling APIs made tool use a product feature months later.