Not an answer, a job

For most of the history of computing, software waited for us. We clicked a button, typed a command or submitted a request, and the machine responded.

That relationship is starting to change. The most important shift in artificial intelligence may not be that machines are getting better at answering questions. It may be that they are becoming capable of pursuing goals.

Give an AI system a destination, access to tools, some memory of what happened before, and enough authority to act, and the interaction looks different. You don't ask it for an answer. You give it a job.

An answer is an output. A job is a process.

An answer can be judged at the moment it is produced. A job unfolds through time. It needs decisions, observations, corrections, and sometimes delegation. It touches an environment. It can succeed, fail, or create consequences nobody could see when the first instruction was given.

That is why the shift is bigger than "chat, but automated". It is a change in the unit of computing. For decades the unit was the program: instructions written in advance to turn inputs into outputs. Then came services, which stayed available, kept state and answered requests. Now we are building systems that can interpret an objective, decide what steps it needs, use tools to take them, look at the results, and keep going until the job is done or a human steps in.

The agent is an assembly

None of this is magic. At its simplest, an agent is a model with a loop around it: reason, act, observe, reason again. A model behind a chat box produces an answer and stops. Put the same model inside an environment where its output can become an action, and it can search a database, run code, open a browser, edit a document, ask another agent for help, check the result and carry on.

So the agent is not just a model. It is a system. Memory gives it continuity, tools give it reach, a runtime gives it somewhere to run, permissions bound its authority, and observability lets people see what it is doing. As models improve, the limit moves outward, from the model to everything around it.

The question that follows

Take a simple request: find me the best flight to San Francisco tomorrow. An agent might work out your airport, check your calendar, search airlines and recommend a flight. A more autonomous one might book it, update your calendar and watch for delays. Each step crosses a boundary, from understanding to action, and from action to ongoing responsibility.

And each boundary raises the same question: who authorized the next action? If the agent hands the search to another agent, was that allowed? If that one calls a third, does the original permission still hold? If something goes wrong, who is responsible?

These aren't philosophical questions. They are infrastructure questions, the kind that only become unavoidable once systems grow large enough for informal assumptions to break. The internet taught us that, and so did the cloud. Agents are starting to teach it again.

This essay is drawn from Chapter 1 of The Agentic Mind, "From Answers to Actions".