Open Data Infrastructure
Data Modeling for Multi-Agent Workflows
How shared entities, task state, handoff records, permissions, and audit context change data modeling for multi-agent workflows.
A single agent can survive with messy context for a while. Multiple agents turn messy context into competing versions of the same business process.
Multiple agents need shared state
The Model Context Protocol describes a way for AI applications and agents to access tools and data sources. OpenLineage and catalog systems give data teams ways to track datasets, jobs, ownership, and data products.
Multi-agent workflows need data models that treat task state, handoffs, permissions, source evidence, and audit context as first-class records. Otherwise, every agent reconstructs the process from partial memory.
The model needs handoff records
A useful model defines shared entities, task instances, agent roles, tool calls, decisions, approvals, denials, and handoff events. It also records which data product or context receipt supported each step.
This is not only for observability. It shapes correctness. If one agent classifies a customer issue and another drafts a response, the handoff should preserve entity identity, policy limits, evidence, and unresolved uncertainty.
Core idea: Multi-agent modeling is workflow modeling with policy and evidence attached.
The ODI modeling pattern
Open Data Infrastructure helps because data products, catalogs, lineage, policy, and provenance already describe the data side of the workflow. The agent workflow model should reference those assets rather than copying their meaning into prompt history.
For adjacent context, read data modeling for tool-calling agents, data modeling for agentic analytics, and context graphs for data product discovery.
What breaks first
- Agents pass natural-language summaries instead of structured handoff records.
- Entity identity changes across steps because each tool uses a different key.
- A later agent acts on data the first agent was allowed to see but the second agent should not receive.
- Audit records capture outputs but not the chain of tool calls and approvals behind them.
Questions to ask
Ask what entities, task states, handoffs, and decisions are modeled explicitly. Ask how permissions and evidence move from one agent to another. Ask which records are replayable after an incident.
If several agents share a job, the data model is the contract that keeps them from improvising the business process.
Sources to start with
These primary sources anchor the technical claims in this guide.
- Model Context Protocol documentation
- Model Context Protocol tools specification
- OpenLineage object model documentation
- W3C PROV overview
- DataHub data products documentation
Multi-agent workflows need a shared data model before they need more autonomy.