What's this about?
DataSpace is a benchmark for "data agents" — LLM agents that do analytics — with an emphasis on *verifiable* answers over messy, heterogeneous workspaces (multiple files, formats, tools), not clean toy tables.
What is a "data agent"?
A data agent is an AI that answers analytical questions by actually operating on your data: loading files, writing and running queries or code, and producing a result. "Verifiable" means the benchmark checks not just whether the final number sounds right but whether the agent's steps actually justify it — closing the gap where an agent confidently returns a wrong-but-plausible answer.
Why does it matter?
"Chat with your data" demos are easy; trustworthy analytics over real, messy workspaces is hard. A benchmark that rewards verifiability pushes the field toward agents you can actually rely on for decisions.