AgentsMeristem

Licensed under Apache-2.0 · Runs local, zero data egress

---

What it is

When data breaks, an agent shouldn't guess. AgentsMeristem detects an anomaly, traces it through the DataHub lineage graph to the real upstream root cause, and writes the finding back so the next person — or the next agent — inherits it. The reasoning core is a fine-tuned 7B model that runs local: your data never leaves the network.

It ships in two forms:

• A capability layer any agent plugs into — as an MCP tool, a DataHub Skill, or a Python SDK.
• A reference product — a root-cause console you can use directly, with a landing page that demos the embed-in-your-agent flow.

Architecture

One engine, many incidents

The same lineage-reasoning core, triggered by different signals:

Scenario	Question it answers	Demo datapack
Freshness incident	"The report went stale — which upstream job is stuck?"	nyc-taxi
Quality collapse	"A column went null — which change introduced it?"	healthcare
Schema drift	"Who breaks if I edit this column?"	ecommerce
PII spread	"Where did this sensitive field flow downstream?"	healthcare
Cost attribution	"What is the source of this costly query?"	extension
Compliance provenance	"Prove this model's training data is clean."	extension

Three vignettes, one capability: 3 a.m., an on-call engineer stares at an empty revenue dashboard — AgentsMeristem walks the lineage and answers in seconds: "the upstream Airflow job failed last night." Tuesday afternoon, an analyst wants to drop a column — it checks forward lineage first: "don't — three dashboards and an ML feature still read it." Quarter-end, a compliance lead needs provenance — it traces back to source and produces an auditable trail.

Open-source contribution

We gave the DataHub-native baseline of this workflow back to the community as a new skill, datahub-triage, in the official skills registry: datahub-project/datahub-skills#32. It localizes a failure using catalog-level signals (lineage, health, assertions, incidents); AgentsMeristem extends it by probing the actual warehouse data for the issues that are invisible in metadata. See packages/agentsmeristem-skill/.

Status

Working end-to-end on real data. Two scenarios run against a live DataHub:

• Freshness (examples/freshness_demo.py) — a stale mart is localized to the staging stage that lags its source, then an incident is written back.
• Data quality (examples/quality_demo.py) — a bad mart is localized to the origin defect, selectively (a billing symptom → the billing defect, not an unrelated one).
• Any agent, via MCP (examples/mcp_client_check.py) — the same tools driven over the MCP protocol.

License

Apache-2.0 © 2026 Vector897