Notes

Short write-ups on agents, storage, and research. New notes are also sent to subscribers on Substack.

Subscribe on Substack ↗RSS

Tracing an LLM agent from its reasoning down to the system callsOct 31, 2026A worked example. I trace a small tool-calling agent on a local model with DFTracer’s agent spans, capture each tool’s system calls in a file of its own, and grow the data 16× twice to see which layer explains what.Capturing and analyzing fine-grained deep learning I/OOct 3, 2026A worked example. I trace a 3D UNet’s training with DFTracer at four data-loader settings and read the traces with DFAnalyzer to answer the question that matters most: how much of the loading does the GPU actually wait for?Using an LLM’s thinking time to prefetch its dataSep 29, 2026An agent spends most of a session waiting on storage, and the seconds it spends reasoning beforehand say which files it is about to read. AgentStage uses that window to stage them.