Programmatic tool calling matches or beats JSON tool calling in 11 of 14 models on BFCL v4

Sources: preprint
Summary: An unreviewed preprint evaluates exposing tools as typed Python stubs invoked through code, with execution and results handled in one agent turn. It reports a 10.6 percent gain over the JSON baseline for the GPT-5.6 family and parity or better in 11 of 14 models on BFCL v4. Under context rot the programmatic form held stable while the JSON baseline degraded 2.3 percent.
Why it matters: The result is a concrete interface choice for anyone building an agent harness, pending independent replication.