llama.cpp Fixes Agent Checkpoint Issue

Bulky-Priority6824 · reddit · 2026-07-13

Version b9978 of llama.cpp fixes a critical checkpoint issue that severely impacted agentic workflows. Previously, every agent turn created a new checkpoint, bypassing the minimum step interval and squeezing the coverage window too tightly. In common tool-calling loops, a single context rollback would cause many checkpoints to be lost simultaneously, forcing the system to reprocess the entire context.

The fix involves gradually evicting checkpoints that are too close to each other in previous tasks, maintaining a wider and longer coverage window. As a result, long-running agent sessions are much less likely to trigger re-computation after a rollback, significantly improving overall efficiency.

Original post →

More from coding & agent

coding & agent channel →