Learn AI /AI Agents and Automation /AI agent
AI agent
- Starter
- AI Agents and Automation
- Policy officers
- Managers & front offices
Also called: agent, agents.
Back to the milestone you came from
In five seconds
What it is
An AI system that can pursue a goal by selecting and carrying out multiple steps, often using software tools, with oversight matched to its risks.
In plain words
An AI agent is a system that tries to complete a task, not just answer once. It can use digital tools, check what happened, and take another step before giving you a result or asking for approval.
An everyday example
Amina needs to get from one side of Kathmandu to the other for two meetings. She gives her travel helper a goal: arrive by 09:00, avoid expensive routes, and include a backup plan in case of delays. The helper checks a map app, looks up train and bus times, compares costs, and writes a short plan with the best route and a second option. If a line is closed, the helper changes the plan before handing it back.
A human travel helper usually has broad real-world judgement and can notice social cues, safety concerns and unusual context in ways current AI systems often cannot. Real AI agents do not reliably understand the situation like a person. They work by processing patterns, tool outputs and instructions, so they may miss important context, use the wrong tool result or take an unsuitable step unless they are carefully designed and supervised.
Remember thisAn AI agent works through a goal, tool calls, observations and next actions. Its authority and human approval gates should match the consequences of what it can do.
This term is contested
The term is used loosely across industry. Some people call any tool-using assistant an agent, while others use the term only for systems with more independent multi-step behaviour, planning or delegated action.
Understand it
Understand it
Why it matters
The term matters because many current AI products are moving from answering questions to taking actions in software. That changes the risk and the governance needs. If a system can search records, send messages, update files or trigger workflows, leaders and staff need to ask different questions: What tools can it access? What approvals are required? What logs exist? When must a human check the result? Understanding this term helps people separate low-risk assistance from higher-risk automation, choose the right controls, and avoid giving a system more authority than it should have.
How an AI agent turns a goal into actions
After viewing this, a reader should be able to explain that an AI agent does not only answer once; it loops through goal, tool use, observation and next action until it reaches a result or asks for help.
01
1. Goal arrives
A person or another system gives the agent a task, such as finding options, drafting something or updating a record, often with limits like time, cost or approval rules.
Draw a box labelled 'Goal' with small tags underneath: 'task', 'constraints', 'permissions'.
02
2. Plan next step
The agent selects a useful next action instead of only producing one reply. It may decide to search, calculate, ask a question or use a connected service.
Draw a box labelled 'Choose next step' with icons for search, question mark and calculator.
03
3. Use tools
The agent sends requests to tools or services such as databases, calendars, maps or email systems through approved connections.
Draw a box labelled 'Use tool' linked to small side icons for database, calendar and web service.
04
4. Observe result
The tool returns information or a status. The agent checks that result and uses it to decide whether it is finished or needs another step.
Draw a box labelled 'Observe result' with a returned document icon and a loop arrow back to step 2.
05
5. Output or hand off
The agent produces a recommendation, draft or completed action, or pauses for human review when confidence, risk or policy requires it.
Draw a final box labelled 'Result or human review' with split icons: check mark and person.
Show a left-to-right sequence of five boxes connected by arrows. Each box represents one stage in the agent process. A thin return arrow from the observation box back to planning shows that the system may repeat steps. Read from left to right, then follow the loop if needed.
Hover or tap each box to reveal a short example: trip planning, meeting scheduling, and record lookup. The loop arrow highlights when the user hovers over steps 2 to 4.
Describe this without the diagram
The diagram is a five-step process shown from left to right. Step 1 is Goal arrives: a user provides a task, constraints and permissions. Step 2 is Plan next step: the system chooses what to do next, such as search or ask a question. Step 3 is Use tools: it connects to approved services like a database or calendar. Step 4 is Observe result: it receives information back and checks whether another step is needed. A loop from step 4 back to step 2 shows repetition. Step 5 is Output or hand off: it either returns a result or pauses for human review.
With reduced motion: Show the same five boxes and arrows as a static diagram, including a visible curved arrow from 'Observe result' back to 'Choose next step', with no animation.
Many AI systems only reply to a prompt. An AI agent goes further. It takes a goal or instruction as input, selects possible next steps, uses tools such as search, calendars or databases, and then produces an action, a draft, an update or a recommendation.
A simple way to picture it is: goal in, steps in the middle, result out. The middle part matters. The system may look at information, call a tool through an application programming interface, or API, which is a standard way for one piece of software to send requests to another, and then use what comes back to select the next step.
For example, instead of only answering “What trains go across the city?”, it might search timetables, compare routes, check delays and draft a travel plan. Some agents can also send messages, create tickets or update a record. That is why they need limits, logging and human oversight.
People often confuse an AI agent with a chatbot or fixed-rule automation. A chatbot mainly holds a conversation. Fixed-rule automation follows predefined rules. An agent can combine conversation, tool use and step-by-step action toward a goal, but its output can still be wrong, unsafe or poorly chosen.
How “A travel helper planning a trip across a city” maps onto the real thing
| In the story | In the AI system |
|---|---|
| Amina's goal with time and cost limits | The user's objective and constraints |
| The helper checking map and timetable apps | The agent using external tools or services |
| Comparing routes and revising after a closure | Step-by-step planning using new observations |
| The final travel plan with a backup option | The agent's output or proposed action |
A worked example
A conference focal point in Nairobi uses an approved agent to prepare a side-event schedule for circulation by close of business.
Starting from: Goal: "Check accepted submissions against the room-booking calendar, identify clashes, propose available alternatives and stop for my approval before changing any booking."
- The agent extracts the goal, source locations, deadline and requirement for human approval.
- It reads the accepted submissions and queries the authorised room-booking calendar.
- For each event, it compares the requested time, room capacity and existing bookings, then records whether the slot is available.
- Where it finds a clash, it searches the calendar for a suitable alternative and adds the proposal to its working list.
- It prepares a draft schedule and a list of unresolved issues, then stops at the approval gate without changing the calendar.
Result: The focal point receives a draft schedule, proposed alternatives for conflicting requests and a short list of issues requiring a decision. No booking is changed until the focal point approves the action.
In a United Nations workplace
A procurement officer in a specialised agency's technical division is preparing the quarterly progress report to a donor. In an approved internal environment, an AI agent retrieves contract milestones and delivery records, checks them against the workplan, flags discrepancies, drafts the procurement section and pauses for review before anything is cleared or shared.
The agent is pursuing a defined goal through several actions: consulting authorised sources, comparing records, updating its working state, drafting text and triggering a human approval gate. It is doing more than producing a reply to a single prompt.
Before you try thisDo not place bids, confidential pricing, evaluation records, vendor banking details, personal data or other procurement-sensitive material in a tool that has not been approved for it. A procurement officer must verify the sources and conclusions. The agent must not contact vendors, alter records, commit funds or submit donor reporting without explicit human approval.
Where you may hear it
- A vendor demo during a procurement meeting
- An ICT governance board reviewing access to internal systems
- A programme review considering automation of quarterly reporting
- A risk review before a tool is allowed to take actions
Avoid confusion
Avoid confusion
Often confused with
Chatbot
A chatbot is a system that takes messages in ordinary language and returns replies in a back-and-forth conversation.
Copilot
A tool that helps a person complete digital tasks by drafting, searching, and sometimes taking limited actions under human direction and review.
Automation
Automation uses a machine or software to carry out set steps with little or no human action after setup.
Side-by-side
Common misconceptions
Common misconception: An AI agent is just a chatbot with a new name.
Correction: Some agents chat, but the key difference is not the conversation style. It is the ability to take multiple steps and use tools toward a goal. A chatbot may only produce text in one turn, while an agent may search, retrieve data, update a system and then report back. Many products mix these features, which is why the terms are often blurred.
Common misconception: If it can act, it can be left alone to run a whole process.
Correction: Action does not mean reliability or judgement. An agent can choose unsuitable steps, use wrong or outdated tool outputs, or continue a flawed plan. The more authority and access it has, the more important guardrails, approvals, audit trails and human oversight become.
Common misconception: An AI agent is the same as ordinary automation.
Correction: Traditional automation usually follows fixed rules, such as 'if field A is empty, send reminder B'. An agent is more flexible because it can adapt its next step based on what it observes. That flexibility can be useful, but it also makes behaviour less predictable and often harder to test.
Quick check
Which example best fits an AI agent rather than a basic chatbot?
The second option shows the key features of an agent: it takes a goal, uses tools, observes results and chooses another step. The first is a one-off answer, the third is static information, and the fourth is fixed-rule automation.
Go deeper
Go deeper
The simple version
An AI agent is a system that tries to complete a task, not just answer once. It can use digital tools, check what happened, and take another step before giving you a result or asking for approval.
The technical version
In technical terms, an AI agent is a goal-directed software system that uses a model, often a large language model, together with tool access, state and control logic to select and execute actions over multiple steps. A typical cycle is: receive goal and constraints, generate or select the next action, call a tool or service, observe the returned result, update its working state, and continue until a stopping condition is met or a human approval gate is triggered. The tools may be reached through APIs, connectors or protocols such as Model Context Protocol in some implementations. The term is used broadly in industry, so boundaries vary: some people include simple tool-calling loops, while others reserve it for systems with more explicit planning, memory or autonomy. In all cases, the important mechanism is closed-loop behaviour: output from one step becomes input to the next step.
Reviewed external resources
Primary sourceModel Context Protocol (opens in a new tab)Model Context Protocol
This official material explains Model Context Protocol and suits readers who want to understand one way agents can connect to tools and data sources.
Deeperlink checked 2026-07-31
EncyclopaediaAI agent (opens in a new tab)Wikipedia
This overview gives a broad, general introduction to the term and suits readers who want a quick starting point with background context.
Intermediatelink checked 2026-07-31
ArticleBuilding effective agents (opens in a new tab)Anthropic
This article explains how AI agents are framed in practice and suits readers who want a clearer product and design perspective.
Intermediatelink checked 2026-07-31
VideoVideo explanations: AI agent (opens in a new tab)YouTube search
These search results lead to varied video explanations of AI agents and suit readers who prefer audiovisual introductions, but the quality and depth will differ by video.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: AI agent (opens in a new tab)Reddit search
These search results show community discussions and examples of how people talk about AI agents, which suits readers who want informal debate rather than a single authoritative explanation.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
An AI agent works through a goal, tool calls, observations and next actions. Its authority and human approval gates should match the consequences of what it can do.