Agents Without Context

AI agents can generate syntactically correct queries but can't validate semantic correctness without structured business context.

The Problem

An agent asked to "join user events with profile data" generates:

SELECT e.*, p.name, p.email
FROM events e
JOIN profiles p ON e.user_id = p.user_id
WHERE e.event_date > CURRENT_DATE - 7

Syntactically perfect. Semantically wrong. It doesn't know:

Valid join key is user_uuid not user_id (legacy field)

Events table has mixed schema versions (v1, v2)

Test events exist in production tables

Schema migration happened mid-quarter

The agent has no way to validate these rules because they exist as tribal knowledge, not queryable metadata. Without context propagation, agents can't distinguish between syntactically valid and semantically correct queries.

What Actually Happens

Hallucinated joins: Agent joins tables on fields that look related but violate business logic (joining on email instead of verified_customer_id). These are common edge cases that break assumptions.

Stale definitions: Agent uses a metric calculation that was deprecated three months ago

Missing filters: Agent doesn't know to exclude test data, pending transactions, or specific edge cases

Wrong aggregation: Agent sums fields that should be averaged or uses simple average where weighted average is required

The query runs successfully. Results are confidently wrong.

What's Missing

Structured context the agent can validate against:

Query: "join user events with profile data"
Context check:
- events.valid_join_key: "user_uuid"
- events.schema_version: "v2"
- events.exclude_test: true
- profiles.compatible_schemas: ["events_v2", "events_v3"]
- migration.completed: "2024-06-15"

Agent generates the query, validates against current rules, adjusts if needed.