Open Data Infrastructure

Data Modeling for Tool-Calling Agents

How entities, actions, constraints, permissions, and failures should shape data models exposed to agents.

A tool-calling agent does not need your entire data model. It needs the part of the model that lets it act without inventing meaning.

The practical problem

OpenAI function calling and tool documentation describe how models can call tools with structured arguments. That shifts data modeling into the action path. The model is no longer just describing entities for storage or analytics. It is describing what an agent can ask a tool to do.

If the exposed model is vague, the agent fills gaps with guesses. It may confuse customer with account, order with invoice, policy with preference, or draft with committed change.

Model entities and actions together

A tool-facing data model should define entities, identifiers, allowed actions, required fields, constraints, permissions, side effects, and failure behavior. The schema should make invalid requests hard to express.

Structured outputs help, but structure alone is not meaning. The model still needs business grain, ownership, examples, policy context, and documented refusal paths.

Core idea: For tool-calling agents, the data model is also an action contract.

The ODI connection

Open Data Infrastructure gives the model a place to get truth from catalogs, contracts, lineage, and policy. The tool schema should not become an isolated copy of the business model. It should reference the same data product definitions humans use.

This helps teams avoid agent-specific shadow semantics. The agent should not learn a different definition of account, region, churn, or freshness because the tool schema was written in a hurry.

What breaks first

  • The tool schema validates JSON shape but not business meaning.
  • The agent can request actions without seeing permissions or side effects.
  • Identifiers are ambiguous, so the agent chooses the wrong entity grain.
  • Failure modes are undocumented, so denials become retries or hallucinated workarounds.

Questions to ask

Ask which entities the tool exposes, which actions are allowed, which constraints are enforced, and which failures are expected. Ask whether the tool schema references catalog and contract definitions or duplicates them.

For adjacent context, read agentic data contracts for tool calls, entity-centric retrieval data modeling, and semantic contracts in ODI.

Sources to start with

These primary sources anchor the technical claims in this guide.

The agent should act through a model that knows what the business object means.

ODI hub Article library Use the scorecard Agentic contracts Entity retrieval