Model, chatbot, agent: what is the difference?

What you will be able to explain

Distinguish a model from the application around it, identify what an observation-action loop adds to tool access, and name safeguards that a system label alone cannot guarantee.

The essential path is the explanation, the five-step activity and the quiz. The presentation and technical notes are optional. You can mark the lesson complete at any time.

Start with the mechanism

A token is a piece of text defined by a tokenizer, not necessarily a word. A distribution assigns probabilities across the possible next tokens. The context is the input available for this step.

Three steps to keep in mind

1
Model
The learned mathematical engine produces scores or outputs from an input.
2
Chatbot
The application prepares context, manages history, applies rules, and presents replies.
3
Agent
An orchestrator can choose a tool, observe the result, and decide what to do next.

Try it yourself

Follow these five steps in order. Everything in the lab runs locally; no real AI model is called.

1 · Predict

Starting from a chatbot with external tools but no observation-action loop, will adding just the loop change the category in this teaching grid? Predict whether it also proves the system is safe.

2 · Manipulate

Press Chatbot to set the baseline: Chat interface and Application instructions on, External tools, External memory and Observation-action loop off. The model is always present.
Turn on only External tools. Read ‘Tool-enabled chatbot’. Keep all other settings fixed; having a tool is not yet the complete loop used by this grid.
Now turn on only Observation-action loop. Compare the category and explanation. Do not turn on External memory: this comparison isolates the loop rather than the whole Tool-enabled agent preset.
Turn the loop off again to check that the category returns. Then switch External memory on by itself: storage does not substitute for the missing observation-action loop.

Language modelCalculates and generates the token continuation.

Always present

Chat interfaceDisplays messages and keeps visible turns.

Application instructionsAdds a role, rules, or format to the context.

External toolsCan search, calculate, read, or modify a system.

External memoryStores and retrieves state outside model parameters.

Observation-action loopObserves an action result before choosing the next step.

Model description

Language model only

The core component transforms input into output, without a conversational interface or added orchestration in this model.

To illustrate an agent in this grid, add:

an application that orchestrates exchanges
tools that can act or inspect a system
a loop that observes a result and chooses the next step

Limits of this classification

There is no single universal technical definition of “agent.”
The lab classifies a prepared architecture; it runs neither a model nor a tool.
A loop guarantees neither safe autonomy, accuracy, nor permission to act.

3 · Observe

The Chatbot preset is a ‘Conversational application’. Adding tools makes it a ‘Tool-enabled chatbot’. Adding just the observation-action loop then gives ‘Agentic architecture’, even while External memory stays off. Removing the loop returns to a tool-enabled chatbot; adding memory alone does not restore the agentic category.

4 · Explain

Explain the difference between having access to a tool and choosing a next step after observing its result. Name one permission limit, one stopping rule and one action that should require human confirmation. Use speech or paper; nothing is collected.

After your own explanationCompare with one possible explanation

An application can expose tools without repeatedly choosing what to do from their results. An orchestrator manages an action-observation loop: propose an action, validate and execute an allowed tool, add its result to context, then continue or stop. Read-only access, a step or time limit and approval before an irreversible action are separate design choices; the ‘agent’ label grants none of them.

5 · Qualify

Open “Compose an AI system”

Check the model in your head

Six short questions, each with an explanation. You can retry or skip the quiz; your best score stays on this device and is shared between languages.

Question 1 of 6

A language model by itself is a complete chatbot interface.

TrueFalse

Keep these three ideas

Go deeper when you need it

The essential path is complete. Open only the resources you need; none are required to finish.

Technical detailWhat the short version leaves out

External memory belongs to the application, not to the model parameters. Selected memories must be retrieved and placed back into context.

The word “agent” has no single universal boundary, so capabilities should be described explicitly.

An orchestrator receives the goal and state, requests an action or answer, validates an allowed tool call, executes it, adds the observation to context and continues or stops. The loop needs bounded steps, timeouts and explicit error handling; generated text alone is not evidence that an action succeeded.

Reading a calendar, adding one appointment and deleting all appointments require different permissions. Minimize access and require human confirmation for important or irreversible actions. A tool result can be wrong because of arguments, data, execution or interpretation.

Audit the model version, instructions and injected data, tool schemas and errors, memory retention, stop rules, permissions, confirmations and action traces. These recommendations concern system design; the local teaching grid does not test or grant any of them.

Reusable teaching resourceOpen the eight-slide presentation

Use this as a recap or a teaching outline. Download the complete Markdown file for reuse without a network request.

The question

Model, chatbot or agent?

Separate the learned engine from the application and orchestration around it. Describe capabilities rather than assuming a product label tells the whole story.

1 / 8

Download the outline (.md)

Sources and further reading

References checked for this English lesson on 11 September 2026. Publication years are listed separately. These explain the mechanisms, not a current ranking of products; real systems and documentation evolve.

Primary sources

Language Models are Few-Shot Learners — Tom B. Brown et al. (2020) (opens a new tab)Describes GPT-3 as an autoregressive model used through textual instructions and examples. Anchors the model component, not a definition of a complete chatbot or agent.
MRKL Systems: A modular, neuro-symbolic architecture that combines large language models, external knowledge sources and discrete reasoning — Ehud Karpas et al. (2022) (opens a new tab)Proposes a modular architecture combining models with external knowledge and reasoning modules. Shows why a model is one component, without making every modular system an autonomous agent.
ReAct: Synergizing Reasoning and Acting in Language Models — Shunyu Yao et al. (2023) (opens a new tab)ICLR 2023 paper, first posted in 2022. Interleaves reasoning traces and actions that obtain observations. A concrete action-observation loop, not a guarantee of reliable or unrestricted autonomy.
Toolformer: Language Models Can Teach Themselves to Use Tools — Timo Schick et al. (2023) (opens a new tab)Trains a model to select API calls and incorporate results into later predictions. Tool access is an added capability; it does not by itself define general autonomy or establish safe permissions.
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks — Patrick Lewis et al. (2020) (opens a new tab)Combines a generator with retrieval from an external document index. Illustrates an added system capability, not a synonym for a chatbot, agent or durable personal memory.

Your local progress

Status: Not started

Progress stays on this device.