When Is an Agent Done?
Imagine a race with no finish line. Runners would just keep going forever, not knowing when to stop! That would be exhausting and pointless. AI agents need a finish line too. They need a way to check: have I reached my goal yet? If yes — stop. If no — keep going. This might sound simple, but it is actually one of the most important things designers have to think about when building an agent. Getting the finish line right makes the difference between an agent that works perfectly and one that runs forever or stops too soon.
The Done Check
After every sense-decide-act loop, a well-designed agent pauses and asks itself one question: have I reached my goal? Scientists call this the done check. It is built right into the agent's instructions. If the answer is YES — the agent stops, signals success, and waits for a new task. If the answer is NO — the agent loops back to sensing and keeps working. Let us look at how different agents do their done check. A robot vacuum: after each loop, it checks — have I covered all reachable floor space? If yes, return to dock. If no, keep cleaning. A smart speaker timer: it checks — has the countdown reached zero? If yes, play the alarm. If no, keep counting. A chess AI: it checks — is the game over (checkmate, draw, opponent resigned)? If yes, stop. If no, keep thinking and making moves.
Every AI agent has a done check built in. After each action, it asks: did I reach my goal? If yes, stop. If no, keep going. That simple check is what turns a loop into a finished job!
What happens after an agent finishes? Some agents reset and wait for a brand-new goal. The robot vacuum returns to its dock and sits there until someone tells it to clean again. Some agents report back. A search agent might send you a message: "I found the best deal — here it is!" Some agents hand off to a human. A medical AI might flag something unusual and then stop, waiting for a doctor to look at what it found. And some agents start a new goal immediately. A smart thermostat that just turned on the heater immediately starts a new goal: check if the temperature is too warm. If it is, turn the heater off. Every designed behavior after the done check was planned by the engineers who built the agent.
Match each agent to what its done check looks like.
Terms
Definitions
Drag terms onto their definitions, or click a term then click a definition to match.
Here is why getting the done check right really matters. Imagine a cleaning robot whose done check was wrong. Instead of checking "have I covered all floor space?" it only checks "have I been running for five minutes?" After five minutes it stops — even if the room is half cleaned! Or imagine the opposite: a robot with no done check at all. It would clean the same floor forever, running its battery down in a circle with no end in sight. A good done check means the agent finishes exactly when the job is truly done — not too early, not too late.
Engineers spend a lot of time making sure the done check matches the real goal. A slightly wrong done check can make an agent useless. Getting it right requires careful thinking about what 'done' really means for that job.
Complete the sentence.
A robot vacuum checks: 'Have I cleaned all reachable floor space?' and if yes, it returns to its dock. What is this check called?
A cleaning robot's done check says 'stop after 5 minutes' instead of 'stop when all floor is clean.' What problem could this cause?
Design the Perfect Done Check
- Think of a simple agent you would like to build. It can be anything: a room tidying robot, a homework reminder, or a pet feeder.
- Write down the agent's goal in one sentence.
- Now design the done check: what question will the agent ask itself after each action to find out if the goal is complete?
- Write the done check as an if-then statement. Example: IF (all food has been placed in the bowl) THEN (stop and wait).
- Also write one wrong done check — one that would cause the agent to stop too early or too late.
- Share both checks with someone and explain why the wrong one would cause problems.