Tell a chatbot to book a table for four on Friday and you'll get a graceful paragraph about how pleasant Friday dinners can be. Tell an agent the same thing and it opens your calendar, checks the restaurant's availability, finds that 7pm is gone, tries 7:45, and comes back with a confirmation. Both may run on the same kind of language model. Only one of them finishes the job.
That difference — answering versus finishing — is the whole idea behind an agent. It's also why agents are both useful and slightly unnerving: software that keeps taking steps toward a goal can be wrong in ways a single reply never is. The good news is that the machinery is easier to understand than the marketing suggests. Four parts, one loop.
The Difference in One Sentence
An agent is software that pursues a goal by repeatedly perceiving its situation, planning a next step, drawing on what it remembers, and acting — then checking the result and going around again.
A chatbot has a turn. An agent has a loop.
Almost everything else follows from that. Loops need memory, because step nine depends on step two. Loops need a way to stop, or they run forever. And loops can fail halfway through in ways that leave real traces: an email already sent, a file already deleted, a payment already made.
Four Parts, Four Boxes
If you sketched an agent on a whiteboard, you'd draw four boxes and two arrows. Each box does one job, and the arrows are where most of the interesting problems live.
Perception: what it can notice
Perception is everything the agent takes in — your instructions, the file it just opened, the error message from its last action, a calendar, a spreadsheet, a web page. It's plumbing as much as intelligence. If the agent cannot see your calendar, it cannot plan around it, no matter how well it writes.
Planning: choosing the next step
Planning turns a vague goal into a short sequence: decide what to do first, which tool to use, and when the job is done. Good planners write brief plans and revise them the moment reality disagrees. Weak ones write long plans and defend them.
Memory: holding on to enough
There are two kinds. Working memory covers the task in front of it — what happened five steps ago. Long-term memory holds preferences and past outcomes. Working memory runs out; that is why longer tasks tend to forget the constraint you gave at the start.
Action: touching the world
Action is where the agent stops talking and does something: sends the message, runs the query, fills the form, updates the record. Actions split into two kinds, and the split matters. Read-only actions gather information. Write actions change something. Every write action is a place where you want a permission check.
The Loop in Slow Motion
Watch the dinner task closely and you can see each part working. Perception reads your message, your calendar, and the booking page. Planning picks a route: check availability, choose a slot, book it. Action fills in the form — and hits a wall, because 7pm is taken.
This is the moment that separates an agent from a chat window. The error isn't an ending; it's new information. Perception reads it. Memory keeps your constraints (Friday, four people, seven o'clock preferred). Planning revises to 7:45. Action completes the booking. Then the loop stops, because the goal is met.
Knowing when to stop matters as much as knowing what to do. An agent that never stops burns your budget and your patience. One that stops too early leaves half-finished work behind. The best ones also know when to stop and ask a person, which is the most underrated skill in the whole cycle.
Why Memory Is Where Agents Stumble
Memory sounds like the easy part. It isn't. Most systems cannot hold an entire long task in view at once, so they keep a window of recent steps and summarize the rest. Summaries lose edges: "no shellfish, and nowhere with a set menu" quietly becomes "dietary preferences." By step twelve, the agent is working from a blurrier version of your request than the one you wrote.
Long-term memory has the same texture. What looks like remembering is usually search: the agent stores notes somewhere and retrieves the few that seem relevant. If the search misses, the knowledge may as well not exist.
There's a fix, and it costs nothing. Put the constraints somewhere the agent can read them, not only in your first message. A short checklist file or a pinned note survives summarization. Restating "four people, Friday, nothing with a set menu" before it acts isn't nagging; it's how you keep the goal intact.
Five Ways Agents Fail
Once you know the four parts, failures stop looking mysterious. They're nearly always one of five things:
- Wrong tools. It can't reach the calendar or the booking page, so it describes what it would do instead of doing it.
- Lost constraints. A detail from early in the task disappears into a
Photo: Pavel Danilyuk / Pexels
