Free GH-700 practice test: 30 questions with answers

Cycle time before and after the change shows whether work is completed faster. Message counts, dashboard colours and repository stars do not measure operational efficiency.

The correction rate shows the quality of the outputs and the hidden effort they create. Speed alone and counts of agents or names do not reveal quality problems.

Clear rules are best handled by deterministic workflows, reserving agents for steps that need judgement. Using agents for every rule adds cost and risk, and manual approval or ignoring the process do not improve efficiency.

Mapping the current process reveals the inputs, outputs and decisions that automation must handle. Building without mapping, naming first and buying an expensive tool skip the analysis.

Quality on a representative sample and measured business outcomes give evidence for expansion. Mentions in meetings, name length and icon counts are not evidence of value or quality.

Reducing backlog age while keeping quality stable shows real improvement without harming outcomes. Closing tickets regardless of quality, message counts and dashboard colours do not measure the goal.

A cost analysis should compare usage and compute costs with the value delivered. Icon colours, name length and desk counts do not reflect agent economics.

A comparable before-and-after sample of the same kind of task gives reliable evidence. Feelings, message counts and selecting only easy tasks bias or fail to answer the question.

Least-privilege scoping limits each agent to its defined tasks, which reduces risk and supports accountability. Administrator access, shared credentials and undefined scope all expand what an agent can do.

Recording inputs, tool calls and reasoning summaries makes the recommendation traceable and explainable. Deleting inputs, showing no details and using random numbers all remove the basis for explanation.

Review and approval steps with output checks catch harmful or inaccurate outputs before customers see them. Public publication, disabled checks and trusting confidence scores alone leave the risk in place.

Data minimization and policy-based protection limit the risk of misuse. Collecting everything, letting the agent share data freely and public folders all breach that principle.

Maximum retry and step limits stop runaway loops and control cost. Unlimited retries, memory allocation and breaks do not limit the agent's behaviour.

The goal, scope and permitted actions, along with the tools and their limits, define what the agent is for and what it may do. Colours, coffee counts and name capitalization are not role definitions.

Repeated errors and actions outside the defined scope both indicate that the agent may be unsafe or unreliable. Steady test performance, on-time routine work and approved tool use are healthy signs.

A required approval step enforces the rule before the action happens. Reminders, wiki notes and trusting memory are not enforcement.

An agent plans multi-step work and uses tools to act toward a goal, usually under human oversight. A single-question responder is closer to a basic chatbot, and the statements about taking actions reverse the distinction.

Triage needs classification, application of rules, and an action in the system, which an agent can perform with tools. Poetry, ignoring rules and fixed answers do not meet the task.

Human-in-the-loop approval before high-impact changes keeps accountability with people. Autonomy without review, approval after the fact and removed logging all weaken control.

An agentic workflow pursues a goal through steps and checks results as it goes. A static web page, a spreadsheet formula and a printed manual do not pursue goals.

Agents suit varied inputs and judgement over unstructured content, where fixed rules break down. Fixed steps with no variation are better handled by simple automation, and tasks with no inputs or no change need no agent.

Clear goals, defined tools and monitoring make agents effective, and because agents can act in systems, their permissions must be controlled. Agents do make mistakes, governance is still needed, and unlimited access is unsafe.

Escalation to a human hands uncertain cases to a person, which keeps quality and accountability. Infinite loops, silent failures and random retries do not provide human oversight.

A pull request keeps human review, tests and checks in front of merges, which is essential for agent-proposed code. The other options remove review or misdescribe the step.

Working on a separate branch and summarizing progress in the issue or pull request keeps the work visible and reviewable. Direct main-branch changes, private laptop files and unrecorded chats reduce transparency.

Issue assignment records who is responsible for the work, which makes ownership clear. Gists, Pages and Sponsors do not assign work.

Requesting clarification prevents wrong actions on ambiguous issues. Closing unclear issues, assigning them at random and ignoring the text all risk poor outcomes.

A failed check should block the merge, be reported, and be fixed by the agent or a person. Merging anyway, deleting tests and hiding failures all undermine quality control.

A knowledgeable person reviewing the documentation through the normal process ensures accuracy before publication. Trusting the agent alone and publishing before customers see errors are not sound practices.

The checks section of a pull request shows the status of automated checks on the latest commit. Sponsors, gists and Pages do not show check results.