presentsUNAIVERSE
The UN AI timeline

    Learn AI /AI Agents and Automation /Agentic AI

    Agentic AI

    Also called: agentic, agentic systems.

    Detail

    In five seconds

    What it is

    A system that can pursue a goal through several tool-using steps, checking results and adjusting its next action with limited human direction.

    In plain words

    A regular AI tool often gives one answer. An agentic one can keep working on a task: use tools, check results, take another step, and ask a person when needed.

    An everyday example

    A programme officer needs to arrange a short work trip. Instead of answering one question, an assistant receives a goal: propose flights within budget, check meeting times, draft an itinerary, and prepare a booking request for approval. The assistant looks at the travel policy, checks flight options, notices one option arrives after the meeting starts, removes it, and tries again. It then drafts the itinerary and pauses for the officer to approve before anything is submitted.

    Where the story stops working

    A human assistant may use broad judgement, background knowledge, and accountability in ways a current AI system does not reliably match. Real agentic systems usually depend on predefined tools, structured permissions, and machine-readable feedback. They do not automatically recognise every exception, and they may continue in the wrong direction unless limits, checks, and stop rules are designed carefully.

    Remember thisAgentic AI acts through a controlled loop: choose a tool, observe the result, update the plan and continue or escalate. That makes approval points, access limits and action logs essential.

    This term is contested

    The boundary of the term is unsettled. Some people use it narrowly for systems that can plan and act over multiple steps with tools and feedback, while others use it more loosely for almost any AI assistant that performs actions. It is best to ask what the system can actually do, with what permissions and oversight.

    Understand it

    Understand it

    Why it matters

    For policy and leadership, the key issue is not whether a system can write well. It is whether it can take actions across tools and systems. That changes the risk profile. Benefits may include speed, consistency, and reduced routine workload. Risks may include hidden errors repeated at scale, unauthorised actions, weak audit trails, and over-trust by staff. When people say they want an “agentic” system, leaders should ask practical questions: What tools can it access? What can it do without approval? What logs are kept? How does it stop? Who is accountable if it acts on bad information? The term matters because governance for a system that can act should be stricter than governance for a system that only drafts text.

    How an agentic system turns a goal into actions

    After viewing, the reader should be able to describe the action loop: goal, plan, tool use, result check, and either continue, stop, or ask a human.

    1. 01

      1. Goal and limits

      The system receives an objective, allowed tools, rules, and stop conditions.

      Draw a rounded rectangle labelled 'Goal and limits'. Inside, show four small tags: objective, tools, rules, stop conditions.

    2. 02

      2. Choose next step

      The model proposes the next action, such as searching, querying, drafting, or asking a clarification question.

      Draw a second box labelled 'Choose next step'. Add a small checklist icon with items: search, query, draft, ask.

    3. 03

      3. Use a tool

      The system calls an external service or software function and gets a result back.

      Draw a third box labelled 'Use a tool'. Show an outward arrow to a small side box labelled 'calendar / database / browser / form', then a return arrow back.

    4. 04

      4. Check result

      The system compares the result with the goal and rules. It looks for success, error, or missing information.

      Draw a fourth box labelled 'Check result'. Add three small status labels beneath: success, error, missing info.

    5. 05

      5. Decide path

      If the task is complete, it returns an output. If not, it loops to another step. If the rules require approval or confidence is low, it pauses for a human.

      Draw a final decision box labelled 'Decide path' with three arrows: to 'Done', back to 'Choose next step', and down to 'Human review'.

    A left-to-right step diagram with five boxes connected by arrows. The reader follows one task from goal to planning, tool use, result checking, and then to one of three outcomes: another step, completion, or human review. Small notes under boxes show what goes in and what comes out at each stage.

    Hovering or tapping each box reveals a short example: for instance, 'search approved supplier list' under tool use and 'budget exceeded' under check result.

    Describe this without the diagram

    The diagram presents agentic AI as a five-step process. First, the system receives a goal, tool permissions, rules, and conditions for stopping. Second, it chooses one next action. Third, it uses a tool such as a database, browser, calendar, or form. Fourth, it checks the returned result against the goal and rules. Fifth, it follows one of three paths: finish and return the result, loop back for another step, or pause for human review. The key idea is that agentic behaviour comes from repeating this observe-and-act cycle, not from one text reply.

    With reduced motion: A static five-box sequence shows the same process without animation. Arrows connect goal and limits to choose next step, then use a tool, then check result, then decide path. The final box has three labelled arrows: done, repeat, or human review.

    Many AI systems only produce one response to one prompt. An agentic system goes further. It can take a goal, choose from available tools, perform several steps, inspect what happened, and continue until it reaches a stopping point or asks for help.

    What goes in can be a goal, rules, and access to tools. A tool might be a calendar service, a database query, a web browser, or software that fills in a form. What happens in the middle is usually a loop: the system plans a next step, acts through a tool, observes the result, and then updates its next step. What comes out is not just text. It may be a completed task, a draft, a set of actions taken, a request for approval, or a log for review.

    This term is often confused with AI agent. They overlap, but they are not always used in exactly the same way. “AI agent” usually names the system itself. “Agentic AI” often emphasises the behaviour or design pattern: goal-directed, multi-step action rather than single-turn answering. In practice, people use the terms loosely, so it helps to ask what actions, tools, limits, and human checks are actually involved.

    It is also often confused with artificial general intelligence. They are different. A system can appear highly agentic in a narrow task, such as updating records across approved systems, without being broadly capable across domains. More action does not mean broad intelligence, and more autonomy increases the need for supervision, permissions, logs, and clear stop conditions.

    How “An office travel organiser with a checklist” maps onto the real thing

    In the storyIn the AI system
    The trip goal and travel policyThe task objective and system rules
    Checking flight options and meeting timesUsing tools to gather information and test whether a step worked
    Removing the bad option and trying againObservation-and-action loop with feedback
    Pausing for approval before submissionHuman oversight and permission gates

    A worked example

    In Amman, a facilities management officer must issue the weekly room-readiness schedule by close of business Thursday, ahead of meetings starting Monday.

    Starting from: Goal: "Schedule repairs for rooms with faulty projectors, use approved technicians within the weekly ceiling, and prepare confirmations for review without sending them."

    1. The system queries the maintenance log, identifies 12 recent projector faults and removes 3 already marked as repaired.
    2. It checks the approved technician roster and available room slots. Only 6 repairs can be completed before Monday within the stated ceiling.
    3. It drafts a schedule for those 6 rooms, prepares unsent confirmation messages and records the remaining 3 faults as unresolved.
    4. It checks whether the goal has been met, detects the unresolved cases and stops. It presents the schedule, drafts, source records and action log to the facilities management officer for a decision.

    Result: The officer receives a proposed schedule for 6 rooms, 6 unsent confirmations and a clearly logged exception list for 3 rooms. No booking is confirmed and no message is sent without approval.

    In a United Nations workplace

    At a multilingual regional hub, a budget officer uses an approved agentic system to help prepare terms of reference for clearance before Friday. The system retrieves the current template, gathers comments from the controlled file, checks the draft against the workplan and budget record, flags inconsistent figures, and produces a revised draft and comment matrix. It stops before circulation and asks the officer to resolve two conflicting instructions.

    The task requires more than drafting. The system selects tools, checks what each returns, retains the status of unresolved points and adjusts its next step. Its stopping rule also matters: conflicting instructions go back to the budget officer rather than being silently reconciled.

    Before you try thisDo not give an unapproved tool access to the file or upload confidential, personal, procurement-sensitive or detailed budget material to it. Connectors should have only the access required. The officer must verify figures, sources and language versions, and the system must not circulate the draft, record clearance or alter an official record without authorised human approval.

    Where you may hear it

    • A vendor demonstration to an ICT governance board
    • A procurement unit reviewing workflow software
    • A regional bureau meeting on automating clearance steps
    • An internal audit discussion on controls for tool-using systems

    Avoid confusion

    Avoid confusion

    Often confused with

    Common misconceptions

    • Agentic means the system can work entirely on its own.

      Usually it means the system can carry out multiple steps with some autonomy inside defined limits. Good designs still use permission controls, stop conditions, logging, and human review for higher-risk steps. The important question is not 'autonomous or not' but 'autonomous to do exactly what, under which rules?'.

    • If a chatbot can call one tool, it is automatically agentic.

      Not necessarily. One tool call inside a single response may still be simple tool use. People often use 'agentic' for systems that can pursue a goal over several steps, inspect results, and adapt the next action. The boundary is not perfectly settled, so teams should describe the actual behaviour rather than rely only on the label.

    • Agentic AI is basically the same as artificial general intelligence.

      No. A system can be very capable at a narrow, structured workflow while still being weak outside that task. Goal-directed behaviour across tools can look impressive, but it does not mean broad, human-level capability across domains.

    Quick check

    Which description best fits agentic AI?

    Go deeper

    Go deeper

    The simple version

    A regular AI tool often gives one answer. An agentic one can keep working on a task: use tools, check results, take another step, and ask a person when needed.

    The technical version

    Agentic AI usually refers to a pattern in which a model is embedded in a control loop that can select actions, call external tools or APIs, observe outputs, update state, and iterate toward a goal under constraints. The mechanism often combines planning, tool selection, function calling or connector use, short-term state, and rules for termination, escalation, or approval. In some systems, orchestration software manages the loop rather than the model alone. The term is used inconsistently: some people reserve it for systems with multi-step autonomy and adaptation, while others use it more broadly for almost any tool-using assistant. For governance, the decisive issue is the operational capability: what actions are possible, what permissions exist, what feedback the system receives, and what human-in-the-loop controls apply.

    Reviewed external resources

    • Primary sourceModel Context ProtocolModel Context Protocol

      This official resource explains Model Context Protocol, which suits technical and governance readers who want to understand one way tools and data sources can be connected to AI systems.

      Deeperlink checked 2026-07-31

    • EncyclopaediaAgentic commerceWikipedia

      This link gives a general encyclopedia-style overview of agentic commerce, which may help readers who want to see one commercial application area of agentic systems.

      Intermediatelink checked 2026-07-31

    • ArticleBuilding effective agentsAnthropic

      This article gives a product-builder view of agents and is suitable for readers who want a practical explanation of how such systems are framed by an AI company.

      Intermediatelink checked 2026-07-31

    • VideoVideo explanations: Agentic AIYouTube search

      This is a YouTube search results page, so it suits readers who prefer video explainers and want to compare several introductions rather than rely on one source.

      search results, not one pagelink checked 2026-07-31

    • DiscussionCommunity discussion: Agentic AIReddit search

      This is a Reddit search results page, so it suits readers who want to see community questions, examples, and disagreements about the term from many posters.

      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

    Agentic AI acts through a controlled loop: choose a tool, observe the result, update the plan and continue or escalate. That makes approval points, access limits and action logs essential.