A new interactive assistant lets you interrogate the AI agent that's supposedly doing your data analysis.
Researchers built MUSE, a meta-agent that sits on top of agentic data science systems - the kind that let you ask an LLM to clean, analyze, and chart a dataset in plain English. It works by breaking an agent's raw execution trace into layered summaries, so a user can zoom from a high-level overview down to the exact step that went wrong. From there, you can click a specific step, ask why the agent did that, and revise it without digging through logs. MUSE also flags steps it judges suspicious and helps translate a user's fix into instructions the underlying agent can actually follow.
Agentic data tools promise to automate away tedious analysis, but they've mostly been black boxes - when one silently drops rows or botches a join, there's been no good way to catch it short of reading raw execution logs. In a 15-person study comparing MUSE against a baseline, users finished tasks faster and reported more confidence they understood and could steer what the agent was doing.
It's a small study and a research prototype, not a shipping product, but it points at the actual bottleneck for agentic tools generally: not capability, trust.