A single timestamp can break your coding agent's prefix cache, Ollama and Together explain

zainhas · x · 2026-11-17

Ollama and Together's joint post explains why coding agents like Claude Code, Codex and Pi stay fast: agents resend the whole conversation each turn, so prefix caching only works if both harness and inference engine do their part. Key lessons: a single timestamp in the system prompt can break the cache; changing reasoning effort wipes it entirely; telling the agent its output was truncated makes it read what it missed.

Related event: Together AI and Ollama Explain How Prefix Caching Keeps Coding Agents Fast(3 posts)→

Original post →

More from coding & agent

coding & agent channel →