Databricks Certified Context Engineer Associate Lessons
45 lessons, one per exam-guide subdomain, in the order the guide teaches them. Every claim is cited to the official documentation.
Domain 1: Foundations of Context Engineering
7 lessons · 16% of the exam
1.1Diagnosing Agent Failures: Harness, Prompt and Structure Levers
Subdomain 1.1: Given a described agent failure, identify the context management technique that would most directly address it.
2 pages · 17 min read
1.2Minimal Tool Sets and Lean Standing Context for Databricks Agents
Subdomain 1.2: Given a Databricks agent design, identify which proactive context management strategies (e.g., minimal tool sets, just-in-time retrieval, tool result scoping) would reduce context window pressure before compaction becomes necessary.
2 pages · 18 min read
1.3Diagnosing Context Poisoning, Distraction, Confusion and Clash from an Agent Trace
Subdomain 1.3: Diagnose context failure modes: context poisoning, context distraction, context confusion, and context clash, given an agent trace.
14 min read
1.4Unity Catalog and Lakebase: Governance and Agent State
Subdomain 1.4: Select the right tool in the Databricks product stack (Unity Catalog, Lakebase, MCP, MLflow 3) for a given scenario.
2 pages · 19 min read
1.5Spotting What Crowds an Agent's Context
Subdomain 1.5: Given a described agent configuration, identify which context elements are consuming disproportionate attention budget and select the change most likely to improve model focus.
2 pages · 16 min read
1.6Reasoning Modes on Databricks: Standard, Extended and Reduced Thinking
Subdomain 1.6: Given a Databricks agent task and its performance requirements, select the appropriate reasoning mode (standard, extended thinking, or reduced thinking) and justify the selection based on token budget constraints and context window impact.
2 pages · 13 min read
1.7Detecting Context Rot in Long-Running Databricks Agents
Subdomain 1.7: Given a Databricks agent operating over an extended interaction, identify the point at which context length is causing measurable degradation in retrieval accuracy or reasoning quality, and select the intervention that restores performance.
2 pages · 17 min read
Domain 2: System Prompt and Instruction Design
4 lessons · 9% of the exam
2.1Genie Space Configuration: Instructions, Example SQL and Trusted Assets
Subdomain 2.1: Given a business domain, select and validate the instructions, sample questions, and trusted SQL assets that together produce a production-ready Genie space.
17 min read
2.2Few-shot example selection under a token budget
Subdomain 2.2: Given a Databricks agent, a set of candidate few-shot examples, and a token budget, justify which examples to include or exclude based on their marginal contribution to agent performance (e.g., covering untested tool paths, demonstrating output structure, handling ambiguous inputs).
13 min read
2.3Revising a Miscalibrated Agent System Prompt
Subdomain 2.3: Given a miscalibrated Databricks agent system prompt and its observed failure pattern, select the targeted revision strategy that resolves the failure with the least increase in token cost and maintenance burden.
15 min read
2.4Comparing Prompt Versions: Quality Deltas vs Token Cost
Subdomain 2.4: Given experiment tracking results comparing two system prompt configurations on a Databricks agent, determine whether the higher-token configuration is justified and identify the specific prompt element driving the cost-performance tradeoff.
2 pages · 18 min read
Domain 3: Knowledge Retrieval and Genie Configuration
9 lessons · 20% of the exam
3.1Diagnosing Genie Agent Accuracy Gaps from Unity Catalog Metadata
Subdomain 3.1: Given an underperforming Databricks agent and its associated Unity Catalog metadata, identify which missing or poorly specified metadata elements are causing the accuracy gap and select the configuration change that will have the highest impact on agent performance.
2 pages · 17 min read
3.2Genie Space Data Selection: Tables, Views and Metric Views
Subdomain 3.2: Select which Unity Catalog objects (managed tables, views, parameterized queries, SQL functions) to curate into a Genie space for a given business domain.
2 pages · 18 min read
3.3Diagnosing AI Search Retrieval Problems with Strategy Evaluation
Subdomain 3.3: Given a Databricks agent with documented retrieval quality problems, diagnose which Databricks AI Search configuration is the root cause and select the remediation that most directly improves the signal quality of retrieved context.
2 pages · 18 min read
3.4RAG Data Pipeline: From a Unity Catalog Corpus to an AI Search Index
Subdomain 3.4: Design a RAG pipeline that retrieves chunks from a Unity Catalog-governed document corpus and injects them into agent context.
2 pages · 18 min read
3.5Chunking Strategies Matched to Document Structure
Subdomain 3.5: Select an appropriate chunking strategy given document structure, embedding model context length, and the types of queries the agent will face.
2 pages · 17 min read
3.6Scoping Context Elements for a Genie Agent Task
Subdomain 3.6: Given a described agent task and available data sources, select the context elements required for the agent to correctly scope and execute the task.
2 pages · 17 min read
3.7Pre-Inference Retrieval: Loading Context Before the LLM Runs
Subdomain 3.7: Distinguish between pre-inference retrieval (embedding-based, up-front loading) and just-in-time agentic retrieval (tool calls, dynamic Delta table queries) and select the appropriate strategy for a given use case.
2 pages · 18 min read
3.8Diagnosing Retrieval Failure Modes from MLflow and AI Search Evaluations
Subdomain 3.8: Identify which retrieval failure mode is occurring by using MLflow eval logs and UC metadata, and select the Unity Catalog governance action that most directly resolves the failure.
2 pages · 19 min read
3.9Certifying Authoritative Data in Unity Catalog for Agent Retrieval
Subdomain 3.9: Given a Databricks agent deployment that will retrieve context from a Unity Catalog environment containing both authoritative and derived data assets, design a governance strategy that constrains the agent's retrieval space to authoritative sources before deployment.
2 pages · 18 min read
Domain 4: Memory Architecture with Lakebase and MLflow
8 lessons · 18% of the exam
4.1Diagnosing Memory Type Mismatches in Databricks Agents
Subdomain 4.1: Given a Databricks agent exhibiting degraded performance because its context is populated from an inappropriate memory source (e.g., retrieving from long-term storage when the information exists in the current session, or relying on session history when the needed context requires cross-session persistence), identify the mismatch between the memory type being used and the information need, and select the correct memory strategy.
2 pages · 17 min read
4.2Durable agent state vs. in-context scratchpad: when an agent needs a store
Subdomain 4.2: Identify when a Delta-backed state object is required over an in-context scratchpad for an agent operating on a multi-step task.
2 pages · 17 min read
4.3Lakebase Memory Retrieval: AI Search vs Structured Queries
Subdomain 4.3: Given an agent retrieving memories persisted in Lakebase, identify whether Databricks AI Search or structured query retrieval is the appropriate mechanism for a specified query type.
17 min read
4.4Setting Up MLflow 3 Evaluations to Compare Agent Context Configurations
Subdomain 4.4: Given MLflow 3 experiment results across multiple agent runs, identify which context configuration produced the most reliable outcomes on a specified task.
2 pages · 20 min read
4.5Static vs. Dynamic Retrieval: Who Decides What Goes Into Context
Subdomain 4.5: Evaluate the tradeoffs of static retrieval vs. dynamic retrieval from Lakebase for a given agent architecture.
2 pages · 18 min read
4.6Over-retrieval and under-retrieval in agent memory
Subdomain 4.6: Diagnose risks of over-retrieval (context pollution) and under-retrieval (missing relevant history) in a memory system.
2 pages · 17 min read
4.7Lakebase-backed agent memory: sessions vs memory stores
Subdomain 4.7: Configure persistent agent memory across sessions using Lakebase-backed durable store.
2 pages · 24 min read
4.8Where to Resolve User Intent in a RAG Agent Pipeline
Subdomain 4.8: Given a Databricks agent pipeline that produces accurate but contextually mismatched responses, identify the pipeline stage where user intent should be resolved before context retrieval.
2 pages · 18 min read
Domain 5: Tool Design, MCP, and Agent Context
6 lessons · 13% of the exam
5.1Progressive Disclosure for MCP Tools on Databricks
Subdomain 5.1: Apply a progressive-disclosure approach to MCP tool access (staged tool discovery with context-efficient execution) to reduce token usage compared to raw tool dumps.
12 min read
5.2Diagnosing Overlapping MCP Tool Descriptions
Subdomain 5.2: Given two MCP tool descriptions that an agent is consistently confusing, identify the overlap in their descriptions that is causing ambiguous tool selection.
14 min read
5.3Progressive Disclosure: Staged Tool Loading to Cut Token Use
Subdomain 5.3: Explain how progressive disclosure can help control how much tool information enters the agent context window at each stage, and why this staged approach reduces token consumption compared to loading full tool schemas upfront.
13 min read
5.4Clearing Raw Tool Outputs from a Deep Agent History
Subdomain 5.4: Given an agent with a deep message history where the context window is approaching capacity, evaluate which raw tool outputs are candidates for clearing.
12 min read
5.5Reading a Unity Catalog Tool's Contract: Names, Descriptions, Schemas
Subdomain 5.5: Given a described agent task and a set of tools registered in Unity Catalog (with names, descriptions, and parameter schemas), identify the most appropriate tool and justify the selection based on functional fit, input/output compatibility, and task requirements.
2 pages · 20 min read
5.6Agent Skills candidates: unloading an overloaded system prompt
Subdomain 5.6: Given a Databricks agent whose system prompt is overloaded with rarely-invoked capability instructions, identify which capabilities are candidates for packaging as Agent Skills and select the loading strategy that minimizes baseline context cost without harming task success rate.
2 pages · 24 min read
Domain 6: Context Compression and Compaction
5 lessons · 11% of the exam
6.1Diagnosing Compaction Failures: What Was Discarded and How to Fix the Prompt
Subdomain 6.1: Given a long running agent task that has been compacted and is now exhibiting downstream coherence failures, identify which category of information was incorrectly discarded during compaction and select the compaction prompt revision that preserves that category without significantly increasing the token cost of the summarized context.
13 min read
6.2Compaction Prompt Tuning: Recall First, Then Precision
Subdomain 6.2: Tune a compaction prompt for a Databricks agent trace: maximize recall first (capture everything relevant), then iterate to improve precision (remove superfluous outputs).
13 min read
6.3Trimming vs. Compaction: Matching the Strategy to Relevance
Subdomain 6.3: Given a long-running Databricks agent task where context window pressure is building, evaluate whether hard-coded trimming heuristics are sufficient for the task's information relevance pattern, or if it requires a more sophisticated compaction approach.
12 min read
6.4Safe Content Removal in Agent Traces During Compaction
Subdomain 6.4: Given a Databricks agent trace, identify which content is safe to remove during compaction without affecting downstream task execution.
12 min read
6.5Aggressive vs. Conservative Compaction: Cost and Fidelity
Subdomain 6.5: Evaluate the tradeoffs between aggressive compaction (lower token cost, risk of losing subtle context) and conservative compaction (higher fidelity, higher cost).
14 min read
Domain 7: Multi-Agent and Long-Horizon Task Design
6 lessons · 13% of the exam
7.1Diagnosing Shared-Context Failures in Multi-Agent Systems
Subdomain 7.1: Diagnose failure modes in multi-agent systems caused by insufficient shared context: inconsistent outputs, conflicting decisions, degraded reliability.
12 min read
7.2Sub-agent dispatch failures: task messages vs agent traces
Subdomain 7.2: Given a multi-agent system where a coordinating agent has decomposed a task and dispatched sub-agents, and the system is now exhibiting downstream failures, diagnose the root cause (i.e., sub-agents receiving individual task messages rather than full agent traces at dispatch time), and select the configuration that resolves the failure without expanding each sub-agent's context window.
11 min read
7.3Preventing multi-agent conflicts through context propagation
Subdomain 7.3: Given a multi-agent system producing conflicting outputs, identify the context propagation change that would most likely prevent the conflict.
12 min read
7.4Reducing Orchestrator Context Load with Sub-Agent Output Design
Subdomain 7.4: Given a multi-agent Databricks workflow where the orchestrating agent is experiencing context saturation, identify the sub-agent output design change that would most reduce orchestrator context load without compromising task coherence.
14 min read
7.5Agent Boundary Placement: Diagnosing Over-Split and Under-Split Designs
Subdomain 7.5: Given a multi-agent system design where agent boundaries have been drawn and the resulting architecture is either exhibiting excessive handoff compression overhead or unmanageable context window growth within individual agents, diagnose which boundary placement failure is occurring.
11 min read
7.6Long-Horizon Agent Failures: Diagnosing Strategy Mismatch
Subdomain 7.6: Given a long-running Databricks agent task with documented performance failures, identify which long-horizon strategy mismatch is causing the observed failure, select the replacement strategy most appropriate for the task's dependency structure, and justify the selection by identifying the specific characteristic that makes the original strategy insufficient.
2 pages · 24 min read