2026-11-14 · America/Los_Angeles · 社区动态 · #10
Letting the model write a tiny DSL instead of JSON tool calls, worth it?
Purpose: looking for real experience before trying this, not promoting anything. Most agent setups have the model return JSON tool calls, one step at a time. An alternative is to give the model a small DSL and let it write a short program for the whole task, which an interpreter then runs and validates. This looks better in several ways: much fewer output tokens, a DSL line is way shorter than JSON one program instead of a round trip per step you can force the grammar while decoding, so even small or quantized models can’t break the format the interpreter catches errors before anything runs.…
热度 44.8 / 100;排名与评分保留该期记录。