Open Data Infrastructure
Agentic Data Reconciliation Workflows
How agentic workflows reconcile proposed writes, source state, table snapshots, human review, and compensating actions.
An agent that can propose a data change also needs a way to prove the change landed where the business expected.
Agents need reconciliation paths
Agentic data workflows often focus on the happy path: read context, call a tool, write a change, and report success. Real systems need reconciliation. The platform has to compare the proposed action, source state, table state, policy decision, human approval, and final outcome.
Without reconciliation, failed writes become awkward mysteries. Partial writes, duplicate actions, stale source reads, and missed approvals all look like normal workflow noise until a customer or analyst finds the gap.
Reconciliation spans several systems
Iceberg snapshots can identify table state before and after a write. Branches can isolate proposed changes. OpenLineage can describe jobs, runs, inputs, and outputs. OPA can record policy decisions. MCP-style tools make action boundaries explicit. None of those pieces replaces reconciliation. Together, they make reconciliation possible.
A good workflow records the proposed write, expected state change, approval decision, execution result, table snapshot, and compensating action if the result does not match the proposal.
Core idea: agentic data is trustworthy when proposed action, approved action, and observed state can be reconciled.
The workflow should leave evidence
For related ODI patterns, read compensating actions for failed writes, agentic data write approval queues, and Iceberg branches for agent experiments.
Reconciliation evidence should be available to the owner of the data product. It should also be available to the system that decides whether the agent can keep acting after failures.
What breaks first
- The agent records the requested action but not the observed table state.
- Human approval is stored outside the write record.
- A failed write triggers a retry without checking source state again.
- Compensating actions are manual and invisible to lineage.
Reconciliation questions
Ask what the agent intended, what was approved, what executed, what changed, and what correction ran if the result was wrong. If the platform cannot answer, the workflow is acting without accounting.
Sources to start with
These primary sources anchor the technical claims in this guide.
- Apache Iceberg table specification
- Apache Iceberg branching and tagging documentation
- OpenLineage object model documentation
- Model Context Protocol tools specification
Agents can act. Data products still need accounting.