Databricks Certified Context Engineer Associate

Learning plan

About 50 hours of focused study, split into 5 phases. Pick your weekly pace and the timeline adjusts.

Estimated finish: 5 weeks

What the exam covers

Source: official exam guide. Weights can change, so check before you start.

Before you start

• Hands-on experience building at least one LLM agent. Databricks recommends six months of platform experience
• Working knowledge of MCP, prompt engineering and few-shot design, and how context length affects model quality
• A Databricks Free Edition workspace for Genie, AI Search, MLflow 3 and agent work

1. Week 1 · 8 h

   Context failure modes and the Databricks stack

   • Diagnose context poisoning, distraction, confusion and clash from an agent trace
   • Apply proactive tactics (minimal tool sets, just-in-time retrieval, tool-result scoping) before compaction becomes necessary
   • Map a scenario to Unity Catalog, Lakebase, MCP or MLflow 3, and choose standard, extended or reduced reasoning for a token budget
   • Spot the point where long context starts degrading retrieval accuracy or reasoning

   Use: Exam guide (official PDF) · Databricks Academy (Generative AI Engineering courses)

2. Weeks 1–3 · 14 h

   Instructions, Genie spaces and retrieval

   • Build a production-ready Genie space by curating tables, views and SQL functions and adding instructions, sample questions and trusted SQL
   • Choose few-shot examples by their marginal value (untested tool paths, output structure, ambiguous inputs) within a token budget
   • Fix a miscalibrated system prompt with the smallest change, and read MLflow experiment results to judge whether a longer prompt pays off
   • Diagnose AI Search problems (chunking, embedding context length, metadata gaps) and restrict retrieval to authoritative Unity Catalog assets
   • Decide between pre-inference retrieval and just-in-time agentic retrieval (tool calls, live Delta queries)

   Hands-on: Create a Genie space over two or three sample tables. Run 10 business questions, then add table comments, a trusted SQL query and two instructions. Re-run the questions and record which answers improved.

   Use: AI/BI Genie documentation · Vector / AI Search documentation

3. Weeks 3–4 · 10 h

   Memory with Lakebase and MLflow 3

   • Match the information need to the right memory type (in-session history, Delta-backed state object or Lakebase long-term store)
   • Choose AI Search or a structured query for memories persisted in Lakebase
   • Spot over-retrieval (context pollution) and under-retrieval (missing history), and compare static with dynamic retrieval
   • Use MLflow 3 runs to compare context configurations across repeated agent runs

   Hands-on: Give an agent cross-session memory by writing user preferences to a Lakebase Postgres table and reading them back at the start of a new session. Log both sessions as MLflow traces and compare them.

   Use: Lakebase documentation · MLflow GenAI documentation

4. Weeks 4–5 · 11 h

   Tools, MCP, compaction and multi-agent design

   • Use progressive disclosure for MCP tools (staged discovery) instead of loading every tool schema upfront
   • Find overlapping tool descriptions that cause wrong tool picks, and move rarely used instructions into Agent Skills
   • Tune a compaction prompt for recall first, then precision, and say which raw tool outputs are safe to clear
   • Compare aggressive and conservative compaction, and choose between trimming heuristics and model-based compaction
   • Fix multi-agent failures by passing full traces at dispatch, slimming sub-agent outputs, or moving agent boundaries

   Hands-on: Register three Unity Catalog functions as tools, two of them with deliberately overlapping descriptions. Trace which tool the agent picks, rewrite the descriptions until it chooses correctly every time, and record the token difference.

   Use: MCP on Databricks · Agent Bricks (Multi-Agent Supervisor)

5. Week 5 · 7 h

   Practice and review

   • Answer the sample questions in the exam guide and tie each one to its objective
   • Rehearse scenario questions on the two biggest domains, Retrieval and Genie (20%) and Memory (18%)
   • Write a one-line fix for each named failure mode (poisoning, distraction, confusion, clash, over-retrieval, compaction loss)

   Use: Databricks AI Prep Guide

You're ready to book when…

• Given an agent trace, you can name the context failure mode and its most direct fix in under a minute
• You have built a Genie space and a Lakebase-backed memory, and can explain when each is the right context source
• You can explain progressive disclosure, compaction recall-vs-precision tuning, and full-trace dispatch to sub-agents without notes

Exam tips

• Almost every objective begins "Given a…", so expect scenario questions that ask for the single most direct fix, not every fix that could work.
• The July 2026 GA guide says "AI Search" where the May 2026 beta guide said "Vector Search". Use the GA guide, and expect both names in older material.