What is an AI agent? Compare it with a chatbot through a document task
Use a meeting-document task to understand an LLM agent’s tool calls and feedback loop, distinguish agents from fixed workflows, and learn which results establish that a task is complete.
Understand the concept
This article focuses on LLM-powered AI agents: systems that choose an action for a goal, obtain external results through tools, and use those results to continue or stop. An agent organizing meeting records might search files, discover that a date is missing, and inspect another record. A basic Q&A interaction generates a response; an agent also dynamically directs execution. One tool call alone does not demonstrate that control. Check the tool records and final result.
The model chooses a step and software executes the tool
A search request produced by a model does not automatically become a successful search. The surrounding software sends it to a tool and returns the tool’s output to the model. The tool might read a file or query a database. A chat response is only one part of this process; the relevant tool must establish what actually happened.
The ReAct paper explored alternating language-model reasoning with external actions, where actions bring back information for later steps. This helps explain an agent’s feedback loop without implying that every product uses the same architecture or understands tasks like a person. Focus on observable tool calls and their results.
Fixed workflows and agents control steps differently
A program that downloads a report each week and totals it with a preset formula follows predetermined steps. Adding an AI summary can still leave it primarily a fixed workflow. An agent might instead notice missing data, inspect another file, and change its next action based on the result. The distinction is how the next step is chosen during execution.
Products do not use the word agent consistently. This article adopts Anthropic’s architectural distinction: predefined code paths are workflows, while agents let the model dynamically direct the process and tool use. A product name alone does not establish its category, and a stable task may not need a more complex agent.
Observe an agent through a meeting-document task
Suppose the task is to extract action items from two meeting records and cite each source. After finding and reading the files, an agent discovers a missing owner. It can inspect another record or mark the item as unresolved. This is an illustrative example; execution records are needed to verify that a particular product behaves this way.
Match each output item to the original record and open the file that was actually created. If a tool failed but the assistant says the file was saved, the task is still incomplete. The table lists evidence you can inspect; a lengthy explanation generated by the model is not proof of success.
| Task step | Evidence to inspect |
|---|---|
| Find the meeting material | File names and readable content returned by the tool |
| Create sourced action items | Items match the source and missing information is labeled |
| Save the final file | The file exists and its contents open correctly |
Set permissions and stopping conditions before use
Reading material and changing an external system require different permissions. A document task can begin with read access, adding write access to the relevant directory when needed. Actions such as sending messages must follow the product’s authorization rules. Having a callable tool does not mean an agent is authorized to perform every action.
Give the task an end condition, such as a completed action-item file with missing information listed. If repeated searches produce nothing new, allow the agent to return its partial result or request more material instead of calling tools indefinitely. This article explains the concept and proposes checks; it does not measure a particular agent’s success rate. The tool-loop article covers further controls.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.