The agentic data stack: a coding agent running dltHub into Snowflake

This is a real dltHub setup, run by Martin Seifert, Data Lead at Pro Juventute. It reframes the stack around how he actually works: the coding agent sits on top and operates everything below it.

Martin drives the loop from OpenCode. It operates through the dltHub AI harness, the agentic primitives that build, run, and fix pipelines, and the dltHub context catalog, which keeps lineage, schema, data quality, governance, run state, and API context in one place the agent reads from and writes to. Those are the two things he calls out: the harness he thinks of as "skills", the catalog as "observability".

Inside OpenCode, Martin splits the work across four specialized subagents: a Data Engineer for ingestion, a Data Modeler for the semantic layer, a Dashboard Developer for data apps, and a Platform Admin for infra, ops, and the build-run-fix fallback that tries a fix and presents it to him before deploying anything.

Underneath, the data plane runs left to right: 25 live pipelines, from Salesforce and Microsoft Teams to Microsoft Planner, Google Tag Manager, an ERP, and OpenCode traces, ingested and orchestrated on dltHub managed infrastructure, landing in Snowflake, and served in Streamlit.