BETTER MODELS: WORSE TOOLS

Summary from the TLDR newsletter

If the newest models get better at solving the task while getting worse at faithfully emitting an alternative tool schema, then the harness needs stronger guarantees somewhere.