· Updated · 4 min read
Context Engineering for Data Agents: Meaning First
Schemas show an agent where data sits. Agreed definitions tell it what the business question means before it chooses an answer.
An answer can run and look plausible while measuring something other than the analyst intended. For leaders deciding how agents use company data, that leaves two credible answers to one request. A schema, prompt, or text-to-SQL model cannot settle business meaning on its own.
What is context engineering?
Context engineering for data agents is the work of assembling versioned business meaning, schemas, lineage, run history and policies around an agent’s goal. It connects one goal to the information needed to interpret, trace and govern how the agent uses company data.
For one goal, the context records the intended purpose, use, users, setting, assumptions and limits. The AI Risk Management Framework from the National Institute of Standards and Technology treats risk management as continuous throughout the AI system lifecycle.
Why does context engineering start with business meaning?
Context engineering starts with business meaning because more context cannot correct an undefined measure or entity. A SpecBridge study tested requests with alternative interpretations. For open requests, the tested models returned both interpretations in only 15–58% of the sets that each model could compute both ways.
The denominator included only sets where that model correctly computed both interpretations when asked separately. A model could calculate each interpretation but still return one without stating the choice. The findings cover 204 hand-validated request sets, four tested models and calculation choices drawn from benchmarks.
An agreed definition must therefore precede prompts, schemas and other context for that measure or entity.
Why isn’t a schema enough?
A schema identifies available data but does not settle what a measure or entity means. Consider the request, “How many manager award winners?” One interpretation counts award records, while another counts distinct managers. One manager can appear in more than one award record, but the second interpretation counts that person once.
Other choices concern missing values and aggregation levels. One result can retain a record with missing related information or drop it. In an aggregation example, weighting individual years equally produced 2.2% annual growth instead of −0.1% when countries received equal weight.
Both interpretations can run and return plausible results. The result alone does not identify the intended business meaning. Once that meaning is agreed, the agent needs the information that surrounds it.
What belongs in a data agent’s context?
Definitions explain the measure or entity. Schemas locate the relevant data. The World Wide Web Consortium’s provenance data model can identify an activity, the information it used and the information it produced.
The following questions show what each part contributes.
| Context | Business question it answers |
|---|---|
| Definitions | What does this term mean? |
| Schemas | Where is the relevant data? |
| Lineage | How is the information traced? |
| Run history | What happened in prior runs? |
| Policies | Which priorities and controls apply? |
The relevant context matches the agent’s stated purpose, use and deployment setting.
Who owns and reviews that context?
The organization remains responsible for the context after the agent starts to use it. NIST calls for accountability structures with appropriate teams and individuals, but it does not assign one universal owner. The organization defines the roles and review frequency for each use.
Review continues after deployment through monitoring, planned periodic reviews and records of performance and impacts.
Leaders can request evidence of:
- Purpose: the intended use and deployment setting.
- Policies: the relevant procedures and controls.
- Responsibility: the people assigned to manage the risks.
- Review: the planned process and its frequency.
- Results: documented performance and impacts after deployment.
These records show whether the organization still understands the context behind an agent’s decisions. They also create the basis for applying agreed meaning in a data platform.
How does dltHub put meaning before code?
dltHub puts meaning before code by treating shared definitions as human-owned information that tools read and apply. dltHub makes this case in our ontology engineering article and its canonical-model article.
dltHub transformation features show how transformations build tables or datasets from loaded data and apply shared business definitions.
The organization keeps the business decision. In the documented example, a sale is a transaction whose status became paid this month. dltHub transformations encode that definition so each related metric uses the same counting rule. dltHub does not decide what counts as a sale.
Meaning comes before agent action
Agreed business meaning comes before the prompt, schema and agent action. Once that meaning is settled, the relevant schema, lineage, run history and policies can support the agent’s goal.
Book a call with dltHub.
What is dltHub?
The agentic data layer on top of any data warehouse: one governed layer and a managed runtime to build, govern, heal and serve data for AI, without growing a platform team.
Was this page helpful?