llama.cpp fixes slowdown in long AI agent workflows
The b9978 change in llama.cpp fixes faulty handling during repeated tool calls by AI agents. Previously, every agent turn created a checkpoint while bypassing the required minimum spacing.
This quickly narrowed the span of earlier context that the covered. If a loop rewound the context, all available could be erased, forcing the model to process the full context again from the beginning.
The fix removes that were created too close together during earlier tasks. This keeps coverage wide and helps long agent sessions continue without costly full reprocessing.
Key points
- b9978 fixes a bug affecting .
- Each agent turn previously created a new checkpoint without normal spacing.
- Rewinding context could erase every checkpoint and trigger full reprocessing.
- The fix removes older that are packed too closely together.
- Long sessions may use less time and computing power as a result.