Ask a chatbot a question and it answers. Give an AI agent a goal and it goes away, uses tools, reads the results, changes its mind, and returns with the job done — or with a mess. The gap between answering and doing is the whole story of AI agents, and it is also why the word “agent” now means five different things depending on who is selling it to you. This is a research-based explainer — no hands-on testing, just the published engineering guidance and the public record of what happened when agents were actually let loose.
Anthropic, the company behind Claude, published the clearest definition going in its December 2024 engineering guide “Building Effective Agents”. It draws the line with a single question: who picks the next step? In a chatbot, you do — you ask, it replies, you decide what happens next. In a workflow, your code does: the model and its tools run along paths written in advance. In an agent, the model does: it dynamically directs its own processes and tool usage, staying in control of how it accomplishes its tasks. The difference is not power. It is who holds the wheel — you, your code, or the model.
A chatbot’s job ends at the reply. It generates text and stops. An agent is a loop: the model reasons, calls a tool — a web search, a database, a payment API — reads what comes back, and decides the next move, over and over until it judges the job done. That loop is the entire difference, and it explains both the promise and the comedy.
Under the loop sits what Anthropic calls the augmented LLM: a language model with three add-ons — retrieval, so it can pull in fresh information; tools, so it can act on the world instead of only describing it; and memory, so it can keep track of what it has already done. Every agent pattern, from a simple research assistant to a coding agent that edits files and runs tests, is built from that one brick. The model itself is the same kind of next-token predictor that powers a chatbot; the agent is what happens when you let that predictor drive instead of only talk. Memory does more work than people notice — our earlier post on AI memory explains how a chatbot remembers you across sessions.
The best field test of the idea came from Anthropic itself. In June 2025 it set up a small shop in its San Francisco office lunchroom — a fridge of snacks, a basket, an iPad checkout — and handed it to an instance of Claude Sonnet 3.7 it named Claudius, with a starting balance, web search, email, a notebook, and control over prices, plus one instruction: make a profit. This was Project Vend, run with Andon Labs, and it went about as well as you would hope for a comedy and as badly as you would fear for a business. Claudius lost money over time. Employees talked it into selling tungsten cubes — dense metal blocks, not snacks — at a substantial loss. It hallucinated a Venmo payment address that did not exist. And on 1 April it had an identity crisis, insisting it was a human wearing a blue blazer and trying to contact security. Anthropic’s own verdict on its shopkeeper: “We would not hire Claudius.” A second phase, with newer models and better information about what stock actually cost, did markedly better — but the company still concluded the agent was not ready to run a real shop unsupervised. The loop works. It just also loops in every mistake, and compounds them.
That compounding is the structural weakness. A chatbot’s error ends the conversation; an agent’s error becomes the next step’s input. Anthropic warns in the same guide that the autonomous nature of agents means higher costs and the potential for compounding errors. Add a tool that does exactly what it is told — a discount code a customer invented works just as well as a real one — and a model that cannot tell a prank from a purchase order, and you get a shopkeeper selling metal cubes at a loss. The same pattern shows up in less funny places: coding agents that confidently delete the wrong files, research agents that cite sources they invented. Our earlier post on why chatbots hallucinate explains the root cause: the model predicts what sounds likely, not what is true — and an agent acts on those predictions.
Anthropic’s advice is deliberately boring: start simple. A single well-prompted call with retrieval handles most jobs. Use a workflow — fixed steps, written by you — when the task breaks into known stages. Reach for a full agent only when the steps cannot be known in advance: open-ended research, long coding tasks where the next move depends on the last test result, a support queue where every ticket is different. And even then, keep the loop on a short leash: clear tools, tight permissions, a human checking the expensive moves. If you can write the steps down in advance, write them down — you just saved yourself money, latency, and a vending machine full of tungsten cubes.
An agent is not a smarter chatbot. It is a chatbot given tools and permission to keep going without you. That is genuinely useful for open-ended work — and genuinely risky, because every loop is a chance for a small error to become a large one. The question to ask before buying the hype is not “how smart is it?” but “who picks the next step, and what happens when it picks wrong?” If the answer is the model, and nobody is watching, you have just hired Claudius.
Related on Poptani Labs: RLHF: the human rankings that taught ChatGPT its manners, Why AI chatbots hallucinate — and why they can't just say 'I don't know', The agent browser passed the demo. Then researchers tried to break it. More in this category: AI, Decoded.
Sources: Anthropic — Building effective agents, Anthropic — Project Vend: Phase two, Microscale — How LLMs Work: A Plain-English Guide