From an answer to a sequence of actions
A chatbot can suggest a packing list. An agent might read an itinerary, look up luggage rules, create a checklist, and save a draft document. “Agent” usually describes a system that uses a model to choose and carry out steps toward a goal, often using tools and feedback. The word is used loosely, so ask what the actual product can do.
An agent is not necessarily a different species of model. The surrounding software provides tools, state, permissions, and a way to continue after each step. A fixed automation follows predefined steps; an agent may decide what to do next within its allowed boundaries.
The useful loop
A simple agent loop is: understand the goal, choose a next step, use a tool, inspect the result, and decide whether to continue. A failed tool call should lead to recovery or a clear explanation. Repeated activity alone is not success.
Give a finish line and a boundary
Try: “Organize these fictional meeting notes into a draft action list. Use only the supplied notes. Mark missing owners and dates. Save a draft for review; do not send messages.” This describes the deliverable and limits the actions.
For your early experiments, keep the task reversible. Review before sending, buying, deleting, publishing, or changing access. Granting an assistant a goal does not mean every possible action is sensible.
A sentence in a prompt is a request, not an access control. Use the app’s real permission settings as well. Start with copies of files and limited tool access; keep passwords and private account details out of practice exercises.
The harness matters
The harness is the software around a model: the tool interfaces, instructions, memory or state, permission checks, retries, and evaluation loop. A model can behave very differently inside different harnesses. When someone says an agent “won” a benchmark, check what software and tools it was allowed to use.