sPTC speeds up agents via speculative tool calling

a1zhang · x · 2026-08-25

Introduced Speculative Programmatic Tool Calling (sPTC), a technique designed to optimize harnesses like RLM and CodeAct. It works by speculating and queuing tool calls (e.g., sub-agent calls) within partially generated code while the root LLM is streaming tokens, effectively hiding latency. The repo includes examples for patching RLM and uses a 'shadowed' REPL namespace to handle input dependencies safely.

Original post →

More from coding & agent

coding & agent channel →