Domain 4 of 6

Tools and Agents

Letting a model act rather than only answer. Function calling and the contract a tool presents, the loop that reasons and acts in turn, what persists between steps, giving a subtask its own clean context, and the failure modes a loop introduces that a single call never had.

Concepts

~18%

Of the exam

Practice questions

Concepts in this domain

01Tool use and function callingThe contract a tool presents to a model, why its description is a prompt and gets read as one, the argument between many narrow tools and few broad ones, and the guard that runs before anything executes.02Agents and their toolsHow an agent reaches the outside world, the four kinds of tool a syllabus usually names, the difference between a tool that knows and a tool that acts, and why the permissions on a tool list decide what a mistake costs.03The agent loopThe cycle in which a model reasons, acts and reads the result before deciding again, ReAct and plan then execute as the two arrangements of it, the token cost of each, and the stop conditions that decide when the cycle ends.04Memory and stateWhat an agent keeps within one run and what survives between runs, the compaction that stops a transcript filling the window, and the difference between a conversation history and a memory somebody designed.05Multiple agents and context isolationGiving a subtask a clean window of its own, the contract a subagent receives, what may cross the boundary in each direction, the token and time arithmetic that makes the split worth it, and what coordination costs in reliability.06Agent failure modesThe five ways an agent fails that a single call never could, the mechanism behind each, and the controls that bound them, which are bounded iteration, a closed tool registry, idempotent calls, approval gates and a full trace.

Try a question from this domain

A car leasing company gives its assistant a tool named get_data, described in full as "Gets data", taking one string parameter called query. The assistant calls it at the wrong moments, and when it needs an account identifier it emits one that is correctly shaped and belongs to no account. Which rewrite addresses both faults?

AA description naming the question the tool answers, the limit on what comes back, what the tool cannot do and the tool that has to run first, together with a parameter description forbidding the model from constructing an identifier.
BA description written for the engineers who maintain the tool, since a precise account of the implementation is what removes the guessing.
CA second tool that validates an identifier before the first one runs, so a constructed identifier is caught by a call rather than by a description.
DA shorter tool set, since wrong calls come from the model choosing among too many tools rather than from what any one of them says.

16 questions on this domain.

One per page, with a worked explanation.

Start the set