DataHub Analytics Agent demo — solving the context problem | DataHub

This is a demo of the open-source DataHub Analytics Agent, walked through by DataHub co-founder and CTO Shirshanka Das. The agent is built around a simple diagnosis: talk-to-data tools get answers wrong because the agent doesn't know what the data means — which table holds the canonical metric, how the business defines a term, whether a dataset is fresh. That's a context problem. In this clip, Shirshanka demos the agent against two real scenarios. In the first, a user asks which sellers have the worst delivery SLA. The agent finds the glossary term the data team already curated, applies that precise definition, and returns a grounded answer. Context quality visibly moves from poor to good within the conversation. In the second, a user asks which product categories are considered sensitive. There's no definition for "sensitive" in the catalog. Rather than guessing, the agent surfaces the gap and offers to close it: a built-in skill reflects on the conversation, proposes specific glossary additions, and writes the updates back to DataHub once confirmed. The next conversation has more to work with. The agent connects three things: DataHub for context (over MCP and the Agent Context Kit), your warehouse for data via standard drivers, and the LLM of your choice. It works against DataHub Core and DataHub Cloud over the same rails. The repo is open-source under Apache 2.0. The writeback loop is what compounds. The agent does useful work today, and the context platform underneath gets richer every time it runs. 🔗 DataHub Open-Source Analytics Agent GitHub: https://github.com/datahub-project/analytics-agent 🔗 Launch blog: https://datahub.com/blog/datahub-analytics-agent/ 🔗 April 2026 Town Hall highlights: https://datahub.com/blog/trusted-context-for-talk-to-data-april-2026-town-hall-highlights/