The Slice From Joe's
Everyone keeps asking what agents are. Everyone loves pizza. So this explainer goes by the slice.
Think about the path of a classic cheese slice from Joe's on Carmine.
Without an agent, you're making the pizza yourself. Sourcing the flour, proofing the dough, pulling the mozzarella, firing the oven, cutting your own slice. Or at the very least, you've got a cold slice in the fridge and you're the one putting it in the oven, watching it, pulling it out, eating it. Every step is on you.
With an agent, you say: "get me a hot cheese slice from Joe's on Carmine, in 40 minutes." Then you go do something else. The agent figures out the route, opens the app or calls the shop, places the order, handles payment, watches the timing, makes sure it shows up at your door warm.
You named the outcome. It handled the moves. That's the whole shift.
A Model With a Goal and a Tool Belt
An agent is a model with a goal, a set of tools, and the freedom to loop until the job is done. Observe what's happening, decide the next step, take it, check the result, repeat.
Anthropic calls this "an LLM using tools in a loop." OpenAI calls it "a system that independently accomplishes tasks on your behalf." Same idea, two front doors.
The loop is the part people miss. A chatbot answers and stops. An agent keeps going. It places the order, sees that Joe's has a line out the door, and adapts without coming back to ask you what to do about it. The goal hasn't changed, so it finds another path to the goal.
That freedom to loop is also why agents feel different to work with. You're not steering every turn. You set the destination and let the thing drive.
What Makes It an Agent and Not a Chatbot
Three behaviors separate the two.
It picks the tools. Which shop, which payment method, which app, which route. You didn't specify that Joe's takes Apple Pay or that the bike lane on Bleecker is faster at 6pm. The agent chose.
It adapts when something breaks. Joe's has a line out the door, so it tries the order-ahead. The order-ahead is down, so it switches to a backup shop that's almost as good. A chatbot would have stopped at the first obstacle and asked you a question.
It keeps going until the goal is met. Not until it has produced one response. Until the slice is at your door, warm, inside the window you gave it.
A chatbot is a smart vending machine. You put a question in, an answer comes out, the transaction ends. An agent is a worker you handed a goal to.
What an Agent Is Not
Three things people mistake for agents, and aren't.
It's not a chatbot that answers questions. Answering is one move. An agent strings many moves together toward an outcome.
It's not a workflow with hard-coded steps. A workflow runs the same five steps every time, in the same order, and falls over the moment reality doesn't match the script. An agent decides the next step based on what just happened.
It's not a magic brain doing everything everywhere at once. It's a worker with a goal and a tool belt, working one loop at a time. The magic is mundane up close.
That last one matters most, because the gap between "magic brain" and "worker with a tool belt" is exactly the gap between a demo that wows you and a tool you can actually trust with a real task.
The Pizza Crawl
Now imagine you don't want one slice. You want a pizza crawl without leaving the couch.
A classic plain from Joe's. A square from Prince Street. A pie from Lucali in Carroll Gardens. A Sicilian from L&B Spumoni Gardens out in Coney Island.
One person running across the city to all four? That's four hours and cold pizza by the time they're back.
So you call your friend who's good at this. Call her the orchestrator. She's in charge of the plan. She sends one runner to Joe's, another to Prince Street, one out to Carroll Gardens for Lucali, and the last to Coney Island for L&B. They all go at the same time. Each one knows exactly which pizzeria and which pie they're carrying.
That's an orchestrator-led multi-agent system. One agent in charge of the plan, a handful of specialist agents doing the legwork in parallel, everything delivered exactly as you asked.
The best part: you didn't call four pizzerias. You talked to your friend, and she handled the rest.
Single agent gets you a slice. Multi-agent gets you the whole crawl.
When You Actually Need the Crawl
The trap is reaching for the crawl when you wanted a slice.
Multi-agent systems shine when the work splits cleanly into independent jobs that can run at the same time. Four pizzerias, four runners, no runner waiting on another. The parallelism is the payoff. Running four coding agents in parallel, each owning one case study end to end, works for the same reason the crawl works: the jobs don't depend on each other, so they don't have to wait in line.
When the work is actually one job, a second agent is just a second person crowding the same kitchen. More coordination, more handoff, more chances for a dropped ticket, and no faster than one good worker would have been.
And the orchestrator isn't mandatory. Sometimes the runners coordinate by sharing one board instead of reporting to a boss. A build of mine, five agents working a room with no dispatcher, did exactly that: every agent read and wrote to one shared space, and the overlap surfaced on its own. No friend making the plan. The board was the plan.
The honest rule: start with one agent. Reach for many only when the work genuinely splits and the single agent is the bottleneck.
Single Slice, Whole Crawl
Strip away the toppings and an agent is simple. A model, a goal, some tools, and the patience to loop until the goal is met. That's the slice.
Put a few of those workers in one kitchen and you get the crawl: parallel specialists, sometimes a friend running the plan, sometimes just a shared board they all read.
The reason the pizza metaphor holds is the same reason the restaurant floor taught me more about shipping than any framework did. Good service was never one person doing everything. It was the right worker, with the right context, sent to the right station, at the same time as everyone else.
Start with the slice. You'll know when you need the crawl, because one runner will be standing in line while three more pies go cold.
What's the part you're still stuck on? That's usually the next thing worth building.