What is an AI agent? Examples, limits and a starter prompt
An AI agent is software that can work toward a goal by choosing steps and using tools, such as search, files or business applications. What it can actually do depends on its connections and permissions. A useful first agent has a narrow task, a checkable result and a clear point at which it hands control back to you.
How is an agent different from a chatbot?
A chatbot can explain how to organize a spreadsheet. An agent connected to an approved spreadsheet tool may also inspect the sheet, propose changes and apply them within the access it has. Product labels are inconsistent, so judge the actual actions and permissions rather than the word “agent.” IBM’s overview explains the role of external tools in modern AI agents.
Source: IBM: What are AI agents?.
| Task | Conversation-only help | An agent with the right tools |
|---|---|---|
| Weekly reporting | Suggest a report outline | Read approved files and draft a report with links |
| Software maintenance | Explain a suspected bug | Inspect a repository, propose a patch and run checks |
| Customer support | Suggest a response | Find the relevant policy and prepare a reply for review |
Choose a first task you can verify
Start with an existing repetitive chore whose correct output you recognize. “Improve the business” leaves too much undefined. “Use these five approved documents to draft a weekly project update, citing each completed milestone” creates a result you can check.
Our suggested task card has six fields: goal, allowed inputs, allowed actions, required output, approval point and stopping condition. This is a planning template, not a claim that every agent product enforces those controls. Check the product’s settings before connecting it.
- Goal: draft one weekly project update.
- Inputs: the five documents you explicitly select.
- Actions: read and draft; no sending or editing source records.
- Output: a table of milestones, owners, dates and supporting links.
- Approval: you check the draft before sharing it.
- Stop: return the draft and list missing or conflicting information.
Try this bounded starter prompt
Draft a weekly project update using only [approved documents]. Return a table with milestone, owner, date, status and source link. Mark unsupported facts as unknown. List conflicts separately. Do not edit source records or send messages. Stop after returning the draft and the questions I need to resolve.
Check the result before expanding the task
Record how long the task takes without the agent. Then time setup, running the agent and reviewing its output. Count corrected facts and missed requirements. Repeat on a few representative examples before connecting more systems. A faster first draft can still cost more time if its sources are difficult to verify.
If the agent cannot complete a step, ask for the missing permission or input and a partial draft. Avoid interpreting a confident completion message as evidence that an external action succeeded. Open the resulting file, record or application to confirm it.
Sources and method
These worksheets and example prompts are suggested workflows, not product test results. This explainer was prepared with AI assistance and links the sources used for background. Product capabilities and access can change.
How we use AI and sourcesKeep up with AI in a few minutes
Get the daily news briefing and useful agent ideas.