A language model can write a supplier email. An agent might find the relevant order, check stock, draft the email and ask a person to approve sending it. That extra sequence changes the work, the cost and the risk.
The short answer: an AI model is the learned component that turns an input into an output. An AI agent is a software system that uses a model to pursue a goal across steps, often choosing tools, observing results and deciding what to do next. In this article, “agent” means a modern LLM-based agent; the broader field of AI also uses the word for other kinds of decision-making systems.
AI agent vs AI model: the difference at a glance
| Question | AI model | LLM-based AI agent |
|---|---|---|
| What is it? | A trained system that maps an input and context to an output, such as text, an image or a prediction. | An application built around one or more models, instructions, tools and a control loop. |
| What does it receive? | The input supplied for one inference call, including any conversation history the application sends. | A goal plus evolving context: earlier outputs, tool results, approvals and state. |
| What does it decide? | What output to generate for that call, possibly including a request to use a tool. | Which step to take next within the permissions and stopping rules the application provides. |
| Can it act on a system? | A model by itself cannot send an email or update a record. It can propose a tool call. | The surrounding software can execute authorised tool calls and record their effects. |
| When does it stop? | When the current response is complete. | When it reaches a completion rule, a step or budget limit, an error or a human approval point. |
| Main failure to watch | A wrong, invented or inappropriate output. | That output plus mistaken actions, repeated errors, tool misuse or unwanted side effects. |
This is an architectural distinction, not a contest between two products. The same model can power a single-turn assistant, a fixed workflow and a more autonomous agent. Anthropic distinguishes predefined workflows from agents that dynamically direct their own tool use; OpenAI’s agent runtime similarly describes a loop of model calls, tool execution and continuation until a stopping point.
What does the model do inside an agent?
Think of the model as a reasoning and generation component. It can interpret a request, extract fields, draft a response or propose a next step. Training gives it learned capabilities; the application supplies current instructions, documents, conversation history and available tools at run time. See our plain-language guide to large language models for how this differs from a database or search engine.
The model does not inherently possess a live inbox, a customer database or permission to pay an invoice. A tool-enabled application exposes specific functions, such as look_up_order or create_draft, then decides whether and how to execute the model’s requested call. In Google Cloud’s architecture guidance, the model is the reasoning engine while the surrounding agent architecture supplies tools, memory and action. This boundary matters: changing the model may improve its judgement, while changing permissions or tool code can change what it is able to do.
Does tool use automatically make a model an agent?
No. A user can ask a chatbot to search once and summarise the result, with every step explicitly initiated by the user. That is tool use, but the application may have little autonomy. At the other end, an agent may choose among tools, react to what each returns and continue until it meets a goal. The label is less useful than asking who chooses the steps, what the software can change and where a person must approve.
One task, three architectures: model, workflow and agent
Suppose an Australian retailer receives a customer message: “My parcel has not arrived. Can you fix it?” The following are illustrative designs, not claims that any particular product behaves this way.
| Design | What happens | Suitable boundary |
|---|---|---|
| One model call | Staff paste the message and approved policy into a model; it drafts a reply. A person checks the order and sends the response. | Useful when staff can supply the facts and the task ends with text. |
| Fixed workflow | Software looks up the order, checks carrier status, feeds those results to the model and routes the draft to staff. The steps are predetermined. | Useful when cases follow a stable process and exceptions can be handed off. |
| Agent | Software receives the goal, chooses whether to inspect an order, query the carrier or request missing information, then revises its plan as results arrive. A refund or external message requires approval. | Useful when cases vary enough that rigid branching becomes brittle, provided permissions and oversight are clear. |
The fixed workflow may be the strongest business choice even when an agent is technically possible. It is easier to predict and test. Anthropic’s engineering guide advises starting with the simplest solution that works and notes that agentic systems often exchange extra time and cost for flexibility. The distinction also explains why a polished conversational interface is not proof of autonomy: it might be a single model call behind a chat window. Our chatbot versus agent comparison goes deeper on that interface question.
How does an AI agent work, step by step?
An agent run usually starts with a user goal, developer instructions and a set of permitted tools. The model proposes an answer or a tool call. The application validates the call, executes it if allowed, returns the observation to the model and repeats. A run ends with a final answer, an error, an approval request or a limit. OpenAI documents this model–tool–model loop; implementations vary in how they store state, handle handoffs and enforce approvals.
- Define the goal and boundary. “Investigate the delayed order and prepare an answer” is narrower than “fix customer problems however you can”.
- Expose only useful tools. Read order status and carrier updates first; leave refunds and sending messages behind approval.
- Observe each result. The agent should distinguish a confirmed carrier update from a model’s guess or an untrusted page.
- Choose the next step. It may ask for an address check, draft a reply or stop because the facts are incomplete.
- Record and review. Keep the tool trail, costs, errors and human decisions so a bad outcome can be investigated.
This surrounding software is sometimes called an agent harness. It carries the permissions, context, loop and stop conditions that a model alone does not provide.
Which should a business build or buy?
Ask what outcome you need before asking for “an agent”. A model-assisted feature is often enough when the output is a draft, classification or summary. A fixed workflow works when the sequence is known. An agent earns its complexity when it must adapt to changing observations across several steps.
| If your task looks like this… | Start with… | Evidence that you need the next level |
|---|---|---|
| Draft a reply from supplied facts | One model call with a clear prompt and review | Staff repeatedly need the software to fetch the same facts. |
| Read records, apply a known rule and draft a result | Fixed workflow with tools and checks | The exceptions require many branches or unpredictable research. |
| Investigate varied cases across approved systems | Agent with narrow tools, budget and approvals | Tests show it completes more useful cases than the workflow at acceptable cost and risk. |
| Send money, delete data or publish externally | Human decision with software assistance | Treat any proposed automation as a separate risk decision with explicit authority and audit. |
Run a small evaluation on real, permissioned examples. Compare completion quality, time, cost, human review effort and errors against a non-agent baseline. For agents, inspect the path, not only the final answer: Anthropic’s agent evaluation guide explains why multi-step behaviour and environment changes matter. Our AI evaluations guide covers how to build a useful test set.
What changes when an agent can take action?
The model can be wrong in both designs. An agent gives a wrong judgement more opportunities to become an external action. A misleading webpage or tool result may also contain instructions that the agent should treat as data, not authority. OWASP describes prompt injection as a risk for LLM applications and excessive agency as a separate risk where functions, permissions or autonomy exceed the task.
Practical controls follow the action boundary: give each tool the least access needed; require approval for irreversible or customer-facing steps; set step and spend limits; log every tool call; test malicious and ambiguous inputs; and define a safe failure path. OpenAI’s agent guidance includes tool guardrails and human review before sensitive side effects. Training a model to be careful helps, but it cannot substitute for application permissions. Anthropic’s research on trustworthy agents makes the same system-level point: model behaviour and the environment around it both affect outcomes.
For Australian organisations, ordinary data handling questions still apply: which records enter the model’s context, where tool results are stored, who may authorise action and how a customer can correct an error. A procurement demo should show the full trail of an agent completing a case, not just its final answer.
Frequently asked questions about AI agents and AI models
Is ChatGPT an AI model or an AI agent?
“ChatGPT” names an application, not one single model or architecture. A chat exchange can involve a model generating a reply; a feature that selects tools and continues across steps can behave more like an agent. Judge the specific feature by its tool use, autonomy and permissions.
Can an AI model use tools without an agent?
Yes. A model can request a tool through an application that performs one controlled lookup and returns the result. The ability to request a tool does not by itself imply a goal-directed loop or broad autonomy.
Does every AI agent use a large language model?
No. “Agent” is a broader AI term. This article compares LLM-based agents with their underlying models because that is the distinction most businesses now mean when discussing generative AI agents.
Is an AI agent more intelligent than its model?
It may complete a task the model cannot complete in one response because it can retrieve facts, use software and correct course. That is a property of the whole system. It does not mean the model’s underlying knowledge or reasoning improved.
Are AI workflows and AI agents the same thing?
Both can combine models and tools. In a fixed workflow, code chooses the route. In an agent, a model has more say over the route within the application’s limits. Real products sit along a spectrum, so ask for the actual decision and approval points.
When should an agent ask a human to approve?
Before a step with material consequences, such as sending a customer message, moving money, changing a record or accessing sensitive information outside the original task. The threshold depends on the organisation’s risk, but approval must happen before the side effect to be meaningful.
What is the simplest way to test whether I need an agent?
Take a sample of actual tasks and build a single-call or fixed-workflow baseline. Add an agent only if it handles the hard exceptions better after measuring quality, cost, review time and unwanted actions. A convincing agent demo is not the same as a reliable operating process.
