Parsing 505 agent-trace datasets on Hugging Face: real user sessions are only ~15%
LysandreJik · x · 2026-10-19
Hugging Face engineer LysandreJik boosted an analysis of the agent traces datasets on the Hub: 505 datasets totaling 17.6GB. The author parsed them all to answer which models, harnesses, tasks and session lengths they involve.
The standout finding: real user usage accounts for only 15% of sessions — the rest are mostly synthetic or eval traces. He notes that digging through agent-trace datasets yields genuinely useful information about how well (or poorly) harnesses and agents are actually being used.
More from coding & agent
- Scrawl: a new tool for effectively critiquing AI agent outputs in design workflows — gregbarbosa · 2026-10-19
- 15 tools every vibe coder should know in 2026, from Claude to Figma — Tegadesigns · 2026-10-19
- Directing agents to work 'within a researcher's interests' becomes a novel prompting trick — MoonL88537 · 2026-10-19
- OpenAI's Dots: Great UI, Broken System — Forgets Tasks, Zero Proactiveness — koltregaskes · 2026-10-19
- Developer spins up a 'manager Claude' just to keep track of his five parallel Claudes — rickasaurus · 2026-10-19
- vLLM 0.32.0 adds structured-decision API reading first-token probabilities — AccBalanced · 2026-10-19