spec-ptc applies speculative decoding thinking to tool calls instead of tokens
GitHub·low signal
alexzhang13/spec-ptc, created 2026-08-21 and at 84 stars, implements speculative programmatic tool calling for harnesses including RLM and CodeAct, and was still being pushed on 2026-08-24. The idea transplants the draft-and-verify pattern from token generation up to the tool-call layer, where the latency cost per wrong guess is far higher than a discarded token. Single-source and early, but it names a specific optimization target that agent harness authors have mostly left alone.