Start now
For teams Agents Integrations Governed Runtime Pricing Enterprise Security Academy Compare Blog
Book a demo Sign in

Module 5 of 5 · 13 min read

Working with AI agents

Everything so far has been about AI that answers. Now the shift to AI that acts. It is the most useful thing AI does at work, and the one that most needs care.

From answering to acting

An AI agent does not just reply to you. It takes actions in real systems on your behalf: it sends the email rather than drafting it, updates the record rather than suggesting the change, works through several steps toward a goal and uses tools along the way. The difference is the difference between "write this email" and "send this email."

That jump is where a lot of the value of AI now lives. It is also where a mistake stops being something you can quietly ignore.

Why acting raises the stakes

A wrong answer is a bad paragraph. You read it, you catch it, you move on. A wrong action is a real event: an email that reached the wrong client, a record changed in the system of truth, a file shared with the wrong people. You cannot un-send it by closing the tab.

Once AI takes the action itself, the question changes. It is no longer only "is this answer any good," it becomes "can we account for what just happened." An agent acting without a record is an action with no author.

What a safe agent looks like

A well-built agent does not act in the dark. It proposes, waits for a human to approve anything that matters, then acts, and leaves a record of the whole thing. Here is that shape.

How a safe AI agent takes an action You make a request. The agent proposes an action. A human approval gate checks it. Only then does the action happen in the real system. Every step is written to a record: who asked, which model, what it did, who approved. You make a request Agent proposes an action Approval a human decides Action in the real system Record who, which model, what
Propose, approve, act, record. The approval gate and the record are what make it safe to use.

Three safeguards to expect

Whether you are using an agent or choosing one, look for three things. If they are missing, that is your warning sign.

  • Approval gates. Anything that writes, sends, or changes something should pause for a human to approve before it happens, not report it afterwards.
  • Least privilege. The agent should have only the access it needs for the task, not the keys to everything. Narrow access limits how far a mistake can travel.
  • A record of what happened. There should be a trail showing who requested the action, which AI model produced it, what it did, and who approved it, joined together, not scattered across systems.

That third one has a name. When the record captures the origin and journey of an action, who asked, which model acted, who approved, and what changed, that is provenance. It is the difference between a pile of log lines and being able to answer, later, for what your AI did.

Your role when an agent acts for you

The rule from module three gets stronger here. An action taken on your behalf is your action. So:

  • Read the approval before you click. The gate only protects you if you actually look. Approving blind is how the wrong email gets sent with your name on it.
  • Do not approve what you do not understand. If a proposed action is unclear, ask or decline. That is the job, not an interruption to it.
  • Know how to stop it. Understand how to pause or shut an agent down, and who to tell, if something looks wrong.
Where this comes from. Approval gates, least privilege, and provenance on every action are exactly what Renly, the company behind this free course, builds into a governed runtime for AI agents. We are telling you the principles so you can expect them from any tool, ours or anyone else's.

Key takeaways

  • An agent does not just answer, it acts in real systems. That is more useful and higher stakes.
  • A wrong action is a real event, so the question becomes: can we account for what happened?
  • Expect three safeguards: approval gates, least privilege, and a joined-up record, which is provenance.
  • An action taken for you is yours. Read the approval, decline what you do not understand, know the stop button.