What you will be able to do
- Name the elements of an agent configuration that take up context: instructions, tool definitions, schema metadata, example queries, conversation history, and retrieved content
- Read a described configuration and point out which element is taking a disproportionate share of context
- Explain why an oversized element hurts focus even when the context limit has not been reached
Key concept
Attention budget (as the Databricks docs frame it) — Everything placed in an agent's context competes for a finite pool of tokens and model focus, so content that is irrelevant, redundant, or unused displaces the content that matters. Improving focus mostly comes from removing or narrowing what you put in, not from adding more.
1.Everything in context competes for the same budget
An exam question for this objective usually describes an agent: its system prompt, the tools or MCP servers it has, the tables and metadata it can see, its example queries, how long its chats run, and what it retrieves. You are asked which of these is taking more than its share and what to change. The phrase "attention budget" does not appear in the Databricks documentation used here. The docs talk about the same problem in terms of token usage, context filtering, and keeping an agent focused. The underlying idea is that every element you load into context costs tokens, and each one competes with the others for the model's focus.
The Genie Agents troubleshooting guide spells out what happens under that pressure. Text instructions and metadata are converted into tokens, and a warning appears as the agent approaches its token limit. Once that warning appears, Genie stops treating every token equally and starts choosing among them.
This is why one oversized element is a problem even before any hard limit. Genie prioritises the tokens it considers most relevant to the question. If hundreds of tokens go to columns nobody queries, the filter may cut the instruction or example the question actually needed. The cost does not show up as an error. It shows up as a weaker answer.
Checkpoint 1 of 4· Check yourself
Which statement best describes how Genie handles a configuration that is approaching its token limit?
The troubleshooting guide says Genie filters by relevance to the question. That is why irrelevant bulk is risky: it competes with relevant content to make it through the filter.
“Genie uses context filtering to prioritize the tokens it considers most relevant to a question.”Source: docs.databricks.com
Sources1
2.Tool definitions: the cost you pay before any work starts
Tools are the element that people most often overlook. Every tool the model could call has to be described to it, and those descriptions sit in context whether or not the tool is ever used. The Genie Code consumption guide is direct about this for MCP: connecting a server brings all of its tools into context, not just the ones a given chat needs.
Checkpoint 2 of 4· Check yourself
An analyst's Genie Code session has five MCP servers connected, but today's chat only uses one of them. What is the effect of the other four?
Every connected server contributes its tools to context, so unused servers are pure overhead. The guide's advice is to turn off servers a chat doesn't need.
“Each connected MCP server adds its tools to what Genie Code carries in context.”Source: docs.databricks.com
The cost is not only tokens. Databricks' AI harness article lists tool overload among the most common harness failures and says a large tool set makes the model's choices worse before it has done anything useful. It also describes a trend away from large collections of narrowly defined tools. The agent design patterns documentation identifies a scale at which a single agent can no longer reasonably hold all of its tools. That is a signal to restructure the agent, not just trim it.
The tool definitions. The system prompt is small and there is no retrieval, so the 60 tool schemas take up most of the context and give the model many near-identical choices. The AI harness article ties this pattern directly to confusion and slower decisions.
3.Instructions, metadata, and examples that repeat themselves
In data agents like Genie, much of the context is metadata: table schemas, column descriptions, sample SQL, and free-text instructions. The troubleshooting guide names each of these as a token driver. It gives the most weight to unnecessary columns, and it also flags descriptions and examples that only repeat information already present.
| Element | What makes it wasteful |
|---|---|
| Table columns | Unnecessary or non-essential columns included from raw tables |
| Column descriptions | Repeating what the column name already says (for example, account_name described as "the name of your account") |
| Example SQL queries | Overlapping or redundant examples, or long examples |
| Text instructions | Unnecessary words, and instructions that don't apply to the agent's questions |
Each of these looks harmless on its own. A column description is good practice. More examples seem like better coverage. A longer instruction seems more precise. They become a problem when they add tokens without adding information the model can use. When you read a scenario, look for duplication (a description that restates the name, two examples that answer the same question) and for irrelevance (columns or instructions unrelated to the questions users actually ask).
Checkpoint 3 of 4· Check yourself
A Genie Agent is near its token limit. Its author wants to keep example SQL useful while cutting tokens. Which approach matches the guidance?
The guide still wants examples for different kinds of questions. Overlapping examples are the waste, so the right move is to keep variety and remove redundancy.
“Include a diverse range of example SQL queries to cover various types of questions, but remove overlapping or redundant examples.”Source: docs.databricks.com
Sources1
4.History and retrieved content that no longer serve the task
Two elements grow while the agent runs instead of being fixed when you configure it: conversation history and retrieved content. A configuration can look lean on paper and still fill up during use.
The consumption guide calls this a "marathon chat" and lists it as a common driver of consumption. History from earlier, unrelated tasks stays in context and makes no contribution to the current one. The Genie Agents troubleshooting guide adds a quality angle: earlier interactions can influence later responses, so starting fresh is also how you get a clean baseline. The AI harness article describes the long-run effect as context rot, where reasoning quality falls as history grows unless something trims or summarizes it.
Checkpoint 4 of 4· Check yourself
Which configuration description matches the documented "marathon chat" driver?
What defines a marathon chat is unrelated tasks sharing one history. A long chat on a single task is a different situation.
“Marathon chats: Continuing one long chat across many unrelated tasks instead of starting a new chat.”Source: docs.databricks.com
Retrieved content works the same way. The RAG quality guidance warns that generation suffers when the context passed to the LLM is missing information or contains information it doesn't need. Both are failures of what you chose to put in. The AI harness article adds that irrelevant retrieval doesn't just waste space. It can lead the model to give wrong answers with confidence. The consumption guide describes a related exploration pattern: asking an agent to read hundreds of assets when it only needs one folder.
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.Tools only cost context when the model actually calls them, so connecting extra MCP servers "just in case" is free.Why is that wrong?
A connected server's tools are carried in context whether or not they are used, so unused servers take budget from every chat.
Covered in Tool definitions: the cost you pay before any work starts
2.A thorough description for every column always helps, even when it just repeats the column name.Why is that wrong?
Descriptions that repeat the column name add tokens without adding information. The guide says to leave them out.
Covered in Instructions, metadata, and examples that repeat themselves
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“Genie uses context filtering to prioritize the tokens it considers most relevant to a question.”
↩︎ Everything in context competes for the same budget“Text instructions and metadata in a Genie Agent are converted into tokens.”
↩︎ Everything in context competes for the same budget“Unnecessary columns in your tables can significantly contribute to token usage.”
↩︎ Instructions, metadata, and examples that repeat themselves“Simplify instructions: Verify that your instructions are clear and concise. Avoid unnecessary words.”
↩︎ Instructions, metadata, and examples that repeat themselves“Previous interactions might influence Genie's responses in any given chat”
↩︎ History and retrieved content that no longer serve the task“Review your instructions and try to focus the agent by removing irrelevant tables or instructions.”
↩︎ Key concept“avoid duplicating information already conveyed by column names.”
↩︎ Exam trap 2“quality may be reduced if important context is filtered out.”
↩︎ Prediction“Include a diverse range of example SQL queries to cover various types of questions, but remove overlapping or redundant examples.”
↩︎ Checkpoint - 2.
“Each connected MCP server adds its tools to what Genie Code carries in context.”
↩︎ Tool definitions: the cost you pay before any work starts“Marathon chats: Continuing one long chat across many unrelated tasks instead of starting a new chat.”
↩︎ History and retrieved content that no longer serve the task“Overly broad exploration: Asking Genie Code to read hundreds of assets when it only needs a specific folder.”
↩︎ History and retrieved content that no longer serve the task“Disable MCP servers you aren't using. Each connected MCP server adds its tools to what Genie Code carries in context.”
↩︎ Exam trap 1 - 3.https://www.databricks.com/blog/ai-harnessSecondary source
“Giving the model too many tools at once increases confusion and slows decision-making before any work begins.”
↩︎ Tool definitions: the cost you pay before any work starts“Developers are moving away from large collections of narrowly defined tools.”
↩︎ Tool definitions: the cost you pay before any work starts“Without a strategy to trim or summarize older context, performance often breaks down on long-running tasks.”
↩︎ History and retrieved content that no longer serve the task“When the harness pulls in the wrong information from memory or search systems, the model may confidently generate incorrect answers.”
↩︎ History and retrieved content that no longer serve the task - 4.
“You have so many tools that fitting them all into one agent's schema is impractical; each agent can own a subset.”
↩︎ Tool definitions: the cost you pay before any work starts - 5.
“It's difficult to generate high-quality RAG output if the context provided to the LLM is missing important information or contains superfluous information.”
↩︎ History and retrieved content that no longer serve the task