Databricks Certified Context Engineer Associate

Not this time. Anthropic names bloated tool sets, which cover too much or leave ambiguous choices between tools, as one of the most common agent failures: if an engineer cannot say which tool fits, the agent cannot either. A small, distinct tool set cuts the tokens spent on schemas in every request and the confusion between them. A larger window still fills with the same noise, earlier compaction treats the symptom and loses detail sooner, and a prompt line does not remove the schemas from the context.

Not this time. The context holds 2 contradictory statements of the same rule, which is context clash, and the model resolves it differently from one question to the next. Databricks' Genie guidance says to keep all instruction types consistent, its example being that if text instructions round to 2 decimals the example SQL must too. Fixing the conflict at its source removes it; outnumbering it with more examples leaves it in the context, and rebuilding the agent is far more than the problem needs.

Not this time. Databricks advises preferring SQL expressions and example SQL over text instructions: SQL expressions define business semantics such as metrics and filters exactly, and text instructions are the last resort for what SQL cannot express. A paragraph leaves Genie to translate the rule into SQL each time, and sample questions show what people ask, not how to compute the answer.

Not this time. Pre-inference retrieval loads context up front, which suits stable reference text but goes stale on hourly data and spends tokens on rows the question never needs. In the just-in-time approach Anthropic describes, the agent holds lightweight identifiers and loads data at runtime with tools, so it reads current figures and only those the task needs. A hybrid, some context up front and more on demand, is also common; a whole inventory in the prompt is the opposite of a minimal context.

Not this time. Databricks' guidance is to keep a Genie Agent focused: include only the tables needed for its questions, aiming for 5 or fewer, prejoin or denormalise related tables in views or metric views, and limit columns. Clear Unity Catalog names and descriptions matter for accuracy. A Genie Agent supports up to 50 tables, views or metric views, so a higher limit is neither available nor the fix, and instructions or samples that patch each join keep the confusing schema in the context.

Not this time. The certified and deprecated system tags in Unity Catalog signal which assets meet your standards and which are outdated, and certified assets show a check mark in the workspace, but a tag does not stop anyone reading a table. Permissions do: an agent can only retrieve from what its identity is granted, so granting it only the certified tables constrains its retrieval space before it runs. A prompt rule can be ignored, and a copy in another workspace is one more copy to go stale.

Not this time. Sessions hold the state of one interaction, such as its transcript; memories hold durable facts, preferences and decisions that should matter in later, unrelated conversations, retrieved by semantic search. A memory pipeline can distil sessions into memories, and the memory stays when the session is cleared. Replaying every past conversation is the over-retrieval that pollutes context, and a preference written into the shared system prompt applies one user's choice to everyone.

Not this time. Anything kept only in the context window is lost when the process ends. Databricks' agent sessions persist an interaction's state in Lakebase so it survives restarts and can be resumed, and a Delta table of processed records does the same for a batch task. A scratchpad, compaction or a bigger window all still live inside the context, so a restart loses them.

Not this time. Agent Skills, the SKILL.md format Databricks also uses, rely on progressive disclosure: a skill's name and description, about 100 tokens, are always loaded, and its instructions and files enter the context only when the task calls for them. The runbooks stay complete for the requests that need them and cost little for the rest. A 2-sentence summary loses the detail, and moving text to the first message costs the same tokens.

Not this time. Anthropic calls tool result clearing one of the safest, lightest forms of compaction: once a tool was called deep in the history and its result used, the agent rarely needs the raw output again. The user's constraints, the system prompt and the result in use are what the rest of the task depends on.

Not this time. Anthropic recommends tuning compaction prompts on complex agent traces: first maximise recall so the summary captures every relevant piece, then iterate on precision to drop superfluous content. Starting short risks the over-aggressive compaction that loses subtle context whose importance only shows later. Simple traces do not exercise what compaction gets wrong, and keeping the last 20 messages is trimming, not compaction.

Not this time. A fixed trimming rule assumes old messages stop mattering, which fails when the earliest turns hold constraints the whole task depends on. Compaction that summarises while keeping decisions and constraints, or structured notes the agent writes outside the context and reads back, keeps them at low cost. A longer trimming window only delays the failure, and never compressing runs into the context limit and its loss of focus.

Not this time. In the sub-agent pattern Anthropic describes, each sub-agent may use tens of thousands of tokens exploring, but returns only a condensed summary, often 1,000 to 2,000 tokens, so detail stays isolated in its context and the lead agent synthesises. Running them in sequence or enlarging the window still pours the same transcripts in, and a single agent puts all of it in one context.