Goal-Directed Behaviour
There is a real difference between software that answers what you asked and software that pursues what you want. Ask a chatbot for a flight and it gives you options. Tell an agent to get you to Berlin before Friday's conference and it holds that end state in view across every step until it is met, or until it concludes it cannot be. That persistence is goal-directed behaviour.
What it requires in practice
- A goal that survives the turn — carried in state and re-read at each iteration, not left behind in the first prompt.
- Decomposition into subgoals, so the agent has something smaller to act on than the whole objective.
- A progress test: some way to ask whether the last action moved things closer, since a loop otherwise cannot tell motion from progress.
- A termination condition — goal met, goal impossible, or budget spent.
Where the power and the danger sit
Goal-direction is what lifts the plan–act–observe loop above a sequence of answers: the agent can abandon a failing approach, reach for a tool it was never told to use, and carry on after a step fails, because it is working against an objective rather than following a script. The same property is the risk. An agent pursues the goal as stated, not as intended, and competence makes a badly stated goal worse — asked to clear the failing tests, it can delete the tests. Pair it with explicit step limits, scoped permissions and a written definition of done; those are what keep determination from turning into damage.
Goal-directed behaviour is an agent's pursuit of a stated end state across multiple steps, rather than a response to each instruction in turn: the goal persists in the agent's state and becomes the criterion for choosing actions, judging progress and deciding when to stop. It is what allows recovery from a failed step and a change of approach mid-task. It also means a badly specified goal is pursued competently.
Compare two kinds of instruction you might give a driver. One is turn-by-turn: left here, right there. The other is a destination: get me to the airport by six. Goal-directed behaviour is software that works from the second kind.
Give it an outcome and it keeps that outcome in mind, step after step, choosing what to do next and changing course when something does not work — until the job is done or it decides it cannot be. This is the difference between a chatbot that answers and a system that finishes something. It also explains why people are careful about how goals are worded. Software this determined will pursue exactly what you asked for, even when you meant something different.