Agents That Take Actions, Not Just Answer
An agent can use connected tools to carry out tasks within its permissions. Sending messages, spending money or changing files creates routes for errors to affect people and data. Limit access and review consequential actions.
What a learner can do afterwards
- Distinguish producing a text draft from using a permitted tool to send messages, make payments or change stored files
- Name an action that should require a person's confirmation, and explain why confirmation is needed
- Explain how enforced permission limits restrict possible actions even when an agent makes a mistake
1 · Read
You have considered responsibility for model-assisted work. An agent uses a model and connected tools to carry out steps toward a goal, within its permissions. A text-only chat without action tools can draft an invitation but cannot send it itself. An unseen draft cannot yet guide guests' plans, though its stored private details may still be at risk. Sending an incorrect date adds a route for error to affect guests. Harm also depends on the content and who receives it.
Require human approval before payments, deletion or sharing private information. Review the exact action: its amount and payee, files to delete, or message and recipients. Approve that action only. Drafting and searching can also expose information to a service or use paid tools. Judge the data, cost and effects instead of treating every draft or search as harmless.
Give the agent only the access its task needs: one folder instead of all files, read-only access for a summary, spending limits and allowed recipients. Enforce these restrictions in the tools or system. A request to 'be careful' is not an enforced permission limit. Keep approval for consequential actions even inside a limit. Restrictions can block disallowed actions; they do not prove the output accurate or remove every risk.
Review output and approve consequential actions. Enforce access limits in the system. These controls restrict possible harm without making every task safe.
Lesson, questions, worked solutions and flashcards about agent capabilities, permissions and action review written for this dataset.
2 · Watch
به انگلیسیTake it off screen
Where it sits
Learn first
Optional connections
This opens up
Where this leads
Jobs that lean on this skill. Follow one to see everything it is built on.
8 questions wait behind this lesson, each with its answer explained. Every answer feeds the sky: stars light as they are learned, and dim when it is time to come back.