The Three Dimensions
When building context for your analytics agent, you need to balance three interconnected performance dimensions:- Reliability — Answer rate and accuracy. Can the agent answer questions correctly?
- Speed — Response time. How fast does the agent respond?
- Costs — Token consumption and query execution costs.
The Approach
Context engineering follows the same methodology as data engineering:- Measure — Track accuracy, response times, and token usage across real-world usage patterns
- Iterate — Identify failure patterns and refine context accordingly
- Optimize — Find the right balance between sufficient context and efficiency
Four Concrete Rules
Rule 1: MECE (Mutually Exclusive, Collectively Exhaustive)
Your context must be:- Collectively Exhaustive — Cover all possible user queries. If a user can ask about it, the context should support it.
- Mutually Exclusive — Each metric has one and only one definition. No conflicting or duplicate information.
Rule 2: Balance Token Costs
Include only relevant schemas, tables, and documentation while eliminating redundancy. This controls costs without sacrificing quality.- Remove unused tables and schemas
- Eliminate duplicate definitions
- Keep context focused on what matters
Rule 3: Minimize Exploration
Provide sufficient upfront documentation to prevent exploratory queries:- Include explicit relationship documentation between tables
- Provide example query patterns
- Document JOIN paths and common aggregations
- Give the agent enough context so it doesn’t need to run discovery queries
Rule 4: Modularity
Organize context into domain-based logical units:- Structure content by business domain (marketing, finance, product, etc.)
- Use hierarchical organization
- Allow agents to access focused information pieces rather than processing everything simultaneously
- Load only relevant context for each query
Next Steps
- Context Engineering Playbook — Step-by-step implementation guide
- Evaluation — How to test and measure your agent’s performance