Launch HN: BitBoard (YC P25) – Analytics Workspace for Agents

Connor and Ambar's startup started out solving a totally different problem: AI agents for administrative grunt work in healthcare. But customers kept dragging them toward something messier — queries scattered across five different sources, spreadsheets nobody could trust, and no single place to see what anyone had already figured out. So they built tooling to fix that, and at some point the tooling became the actual company. That company is now BitBoard, and today it's launching dashboards built for people and AI agents to edit side by side.

The pitch is pretty pointed at two failure modes. AI coding assistants and chat tools treat every data analysis as throwaway — ask a question, get an answer, and then that insight vanishes the moment the chat closes. Meanwhile legacy BI platforms were built for humans clicking around, and now they're just duct-taping a chatbot on top, which doesn't give agents any real ability to act or verify anything. Connor and Ambar say the deeper issue is that software today can make a business far more transparent and measurable than old-school BI ever managed, but neither camp is architected to take advantage of that.

Their fix is to give humans and agents the same underlying data — same entities, same defined metrics, same semantic model — but different tools layered on top depending on who's using it. Dashboards start as raw SQL or code and can escalate into full embedded apps as needed. Every number that comes back is traceable, with provenance attached, and querying the same thing twice gives you the same answer instead of a different hallucinated one. That consistency piece matters more than it sounds: without it, nobody, human or agent, can actually trust the shared workspace.

Under the hood they're running DuckDB and Apache Arrow for columnar analysis, plus what they call grounding and verification infrastructure, along with containers and traces meant to support agents running long jobs rather than one-off queries. The bigger ambition here is autonomous agents that live inside a company's data over time — noticing a metric drifting or a funnel breaking, investigating why, and producing datasets and dashboards a human can review and approve. Their philosophy: let an LLM's judgment spot the anomaly, then hand the actual fix to deterministic, checkable code rather than trusting the model to freelance it.

BitBoard is live now at app.bitboard.work, asking only for an email to get started.