What is an AI agent?

Answering versus acting

An assistant that only answers can be wrong, and you notice. An agent that acts can be wrong in the world: a reply sent to the wrong person, a meeting moved on top of another, a purchase you did not mean. The difference is tools. An agent is a language model that can use your mail, your calendar, a browser or a way to pay.

What agents do well today

• Small errands with a clear end: reply, remind, move, find.
• Watching for something: a mail from one person, a price, a flight.
• Summing up: your day, a long thread, what changed since yesterday.

Where they go wrong

• Misunderstood requests, especially by voice and in a noisy place.
• Confident mistakes: a model can report that it did something it did not do.
• Orders hidden in what it reads: a web page or an email can try to tell the agent what to do. This is called prompt injection.
• Too much access: an agent that can do everything can also undo everything.

Five rules a personal agent should keep

1. Ask before acting. Sending, moving, deleting or paying waits for your yes.
2. Say what it will do, in one line, before it does it.
3. Treat what it reads as information, not as orders.
4. Use the least access it needs, and let you take it back.
5. Keep a record you can read: what it did, and when.

How Lychnos does it

Lychnos proposes; you press once to agree. Sending a reply, moving a meeting or deleting anything waits for that press. In our robot work the same idea runs deeper: the AI proposes, and tested code decides. More in a personal AI agent that asks first and the AI proposes, tested code decides.

Going further: agent vs chatbot vs assistant, how agents get tricked by what they read (prompt injection), and what the EU AI Act requires.

One of the first jobs people hand to an agent is the inbox: what an AI email assistant can and cannot do.