Making a Model Do Things
A model in a loop, with tools and a goal
Scroll
A model in a loop, with tools and a goal
A chatbot answers. You send a message, the model writes a reply, and nothing more happens until you write again. An agent is given a goal instead, and keeps working until it reaches the goal or gives up.
Underneath is the same model, run in a loop. Each time round, it looks at the goal and everything that has happened so far and picks one next step, usually a call to a tool: search something, read a file, run some code. The application runs the tool and sends the result back, and the model picks again. The loop ends when the model says it is done, or when a limit on steps, time or cost cuts it off.
Both the usefulness and the risk come from that loop. An agent can finish a job that takes twenty steps, but every step is a chance to go wrong, and an early mistake carries into everything after it. Agents also act: they can send the email or change the file, where a chatbot could only describe doing so. For any agent, ask which tools it can use, what stops it, and who checks its work. The pages on tool calling, ReAct and agent memory cover those parts one at a time.
Agents