Blog
Agents on migrated pipelines: where to start
By Ted Ghose, Co-founder and CTO · Published · 4 minutes read
The short answer
On regulated pipelines, start small. Let agents check status, retrieve lineage and inspect outputs first; these cannot change a result. Then add one write operation, rerunning a failed step on a non-regulated pipeline, with an approval and a log, and reconcile the output against the last good run. Widen only after a month of clean logs.
Agents are moving into data operations: rerunning failed loads, checking freshness, explaining where a number came from. On regulated pipelines, start small.
Where should agents start on a migrated pipeline?
Start with read operations. Let agents check status, retrieve lineage and inspect outputs. These cannot change a result.
Then add one write operation, with an approval. The first is usually rerunning a failed step on a non-regulated pipeline. Require an approval, log who approved it, and reconcile the output against the last good run. Widen only after a month of clean logs.
What makes a migrated pipeline ready for this?
Pipelines Novuz delivers can be exposed as MCP tools on Databricks, Snowflake or Fabric, under your access, approvals and audit logging. The conversion that built them is not agentic.
Next question: who is liable when an agent reruns a step?
Key terms
- Agent-ready pipeline
- a pipeline that exposes its operations as tools an AI agent can discover and call, under the owner's access controls, approvals and audit logging.
- Idempotent
- a step that, run twice on the same inputs, changes nothing it should not the second time.
FAQ
Frequently asked questions
Our AI program needs agent access to pipelines that were never designed for it. What now?
Do not hand agents database credentials. Expose named operations instead: run, status, lineage, inspect, rerun. Put each behind your identity provider and an approval where the pipeline is regulated. If the pipelines are being migrated, build this in during the move; it is far cheaper than retrofitting it afterwards.
Who is liable when an agent reruns a step and breaks a regulated report?
The same party that is liable when a person reruns it: the pipeline's owner. That is why reruns of regulated steps should require an approval and leave a log. Contract terms on liability are set in each agreement.
Has anyone run agents on migrated pipelines in production?
Ask every vendor for the operations their delivered pipelines expose, under whose permissions, and for the logs. Novuz delivers pipelines that can expose five operations as MCP tools; the First Domain is where the three readiness tests run before any agent touches production.
Sources
- Novuz, what agents can do on your pipelines, and who approves it (novuz.com/agentically-primed/governance)
- Model Context Protocol, official documentation (modelcontextprotocol.io)
About the author
Ted Ghose is co-founder and CTO of Novuz Inc., the deterministic data warehouse modernization company based in Danville, California. He built the Novuz conversion engine and was previously a Cisco principal engineer with more than 15 patents in cloud and networking.
Next step
The First Domain is where the three readiness tests run: idempotence, reconciliation after a rerun, and permissions. One domain, converted and proven in 5 to 6 weeks, with the equivalence guarantee in writing.
Last reviewed: