AI agents explained
Short answer
An AI agent is a software system in which an AI model decides the next step and which tools to use in order to reach a goal a person set. A chatbot answers messages, and a workflow follows a path people wrote in advance. An agent sits between the goal and the path and chooses the steps itself, so it needs firmer limits, permissions and logs.
On this page
Key facts
- What it has
- A goal, an AI model that decides, tools it can call, and instructions with limits
- Who chooses the steps
- The model, at run time, not a person who wrote the path in advance
- Level of autonomy
- Ranges from suggesting an action to carrying it out alone; the business sets the level
- Oversight
- Permissions, logs, retry limits, and a person for high-risk or irreversible actions
- Not an agent
- A single reply from a chatbot, a classifier, or a fixed workflow with one AI step
How an AI agent works
Most agents follow the same loop. The names differ between vendors, but the stages are the same.
- Receives a goal. A person or another system states what should be achieved, for example “find out whether this customer’s order shipped and tell them.”
- Plans steps. The model decides what it needs to do first, based on the goal and the instructions it was given.
- Calls tools. A tool is anything the agent is allowed to use: a search, a database lookup, a calendar, a spreadsheet, an email or a messaging channel. Tools connect through an API where one exists.
- Checks the result. The model reads what the tool returned and decides whether the goal is met, whether to try another step, or whether to stop.
- Reports or hands over. The agent gives the answer, completes the action, or passes the case to a person when it cannot go on.
OpenAI’s guide to building agents describes the same pattern. It says an agent “leverages an LLM to manage workflow execution and make decisions,” recognises when a workflow is complete, and “can proactively correct its actions if needed.” If it fails, it can halt and “transfer control back to the user.” It also says the agent selects among tools “always operating within clearly defined guardrails” (OpenAI, A practical guide to building agents, undated PDF).
The parts of an agent
The same guide says an agent consists of three core components: a model, tools and instructions. The model is the large language model that does the reasoning. The tools are the functions or APIs it may call. The instructions are written guidelines that define how it should behave. Many agents add memory, so that the system keeps facts from earlier steps or earlier conversations.
Levels of autonomy
Autonomy is a setting, not a fixed property. A business can choose any of these, and many start at the first.
- The agent proposes an action and a person approves it.
- The agent acts alone on cases that meet clear conditions and passes the rest to a person.
- The agent acts alone across a wider set of tasks, with every action logged.
The human in the loop entry covers where a person steps in.
Definitions from the model providers and standards bodies
Definitions of the word “agent” vary, and the differences matter when a product is described as agentic.
| Source | What it says | Date |
|---|---|---|
| Anthropic | “Workflows are systems where LLMs and tools are orchestrated through predefined code paths.” Agents “are systems where LLMs dynamically direct their own processes and tool usage.” | 19 December 2024 |
| OpenAI | “Agents are systems that independently accomplish tasks on your behalf.” Applications that use a model without letting it control the workflow, such as simple chatbots, single-turn calls or sentiment classifiers, are not agents. | Undated guide, read 12 October 2026 |
| NIST (CAISI) | “AI agent systems are capable of planning and taking autonomous actions that impact real-world systems or environments.” | 12 January 2026 |
Swipe the table sideways to see all columns.
Sources: Anthropic, Building effective agents; OpenAI guide; NIST, CAISI request for information on securing AI agent systems.
The three agree on the core idea: the model, not a fixed script, decides the next action. They differ in how much weight they give to tools and to autonomy. This site uses the shorter test below.
If a person wrote the order of steps, it is a workflow. If the model chooses the order of steps, it is an agent.
Agent, chatbot or workflow
| Chatbot | Workflow with an AI step | AI agent | |
|---|---|---|---|
| Input | A message in a chat window | A trigger such as a new form, email or payment | A goal, plus context and instructions |
| Who decides the steps | The conversation design or the model’s reply | People, in advance | The model, as it goes |
| Actions it can take | Reply, and sometimes look something up | Fixed actions on every run | Any action among the tools it has been given |
| Predictability | Medium to high for scripted bots, lower for model-based ones | High | Lower: two runs of the same goal can take different routes |
| Typical failure | A wrong or off-topic reply | A step breaks and the run stops | A wrong action, or a long chain of small errors |
| Cost to run | Low | Low to moderate | Higher, because the model is called many times per goal |
Swipe the table sideways to see all columns.
Anthropic’s article makes the trade-off explicit. It says workflows “offer predictability and consistency for well-defined tasks,” while agents are better when “flexibility and model-driven decision-making are needed at scale.” It also says agentic systems “often trade latency and cost for better task performance,” and that autonomy brings “the potential for compounding errors.” For many applications, it adds, optimising single model calls with retrieval and examples is usually enough.
OpenAI’s guide suggests agents suit work where fixed rules fall short: decisions that need judgement, rule sets that have become too tangled to maintain, and tasks that depend on unstructured data such as documents or free-text messages. It adds that if a use case does not meet these criteria, “a deterministic solution may suffice.”
For the related terms, see what AI automation is, what an AI chatbot is, AI assistants and conversational AI.
Types of AI agents
The classic textbook types
The standard university text on the subject, Stuart Russell and Peter Norvig’s Artificial Intelligence: A Modern Approach (4th US edition, Chapter 2, “Intelligent Agents”), groups agents by how they decide. The categories predate language models and still describe the design choices (AIMA, Berkeley).
| Type | How it decides | Business-sized illustration |
|---|---|---|
| Simple reflex | Reacts to the current input with a fixed rule | A rule that tags any message containing “refund” |
| Model-based reflex | Keeps a picture of the situation over time | A booking agent that remembers which slots are already held |
| Goal-based | Chooses actions that move toward a stated goal | An agent that works through steps until an appointment is confirmed |
| Utility-based | Weighs options by how good each outcome is | A scheduler that balances staff hours, travel time and customer preference |
| Learning | Improves its behaviour from experience | A system that adjusts how it ranks leads after seeing which ones closed |
Swipe the table sideways to see all columns.
Types seen in business tools today
Current products tend to fall into groups by task rather than by textbook category.
- Customer-conversation agents answer questions and look up orders or bookings, then hand over to staff.
- Research and summary agents gather material from approved sources and draft a short answer.
- Operations agents work inside business software: sorting a queue, updating records, preparing a report.
- Coding and computer-use agents operate software on a person’s behalf. OpenAI’s guide notes that for legacy systems without APIs, agents can rely on computer-use models to interact with the software directly.
Examples in business
These are illustrations of the pattern, not reports of real deployments.
- Inquiry handling for a clinic. A patient messages a Facebook page asking for a weekday evening slot. The agent checks the open slots in the booking system, offers two, and books the one the patient picks. It passes questions about symptoms or treatment advice straight to staff. Health questions are outside what such an agent should answer.
- Broker lead follow-up. A property inquiry arrives through a web form. The agent reads it, looks up matching listings, drafts a reply for the broker to approve and adds the lead to the CRM with a note.
- Online seller order questions. A buyer asks on Messenger where an order is. The agent looks up the courier status and replies. If the status shows a failed delivery, it flags the case for the seller.
- Scheduling. An agent proposes meeting times across several calendars and sends invitations once a person confirms.
A published example from a Philippine company
PLDT told UiPath that it runs three internal AI systems on UiPath’s platform. One of them, ERICA, supports annual risk reviews by synthesising information, scoring risks and recommending treatments, and another, KAI, retrieves information for proposals and customer support from sources PLDT has approved. The figures in the announcement are PLDT’s own estimates. PLDT did not disclose how it measured them, and no independent audit accompanied them (TechNode Global, 10 September 2026, citing a UiPath announcement of the same date). This article quotes no figures from it.
Agent builders
People who do not write code can assemble agents with several kinds of tools. The categories matter more than any one name, because products change.
- Workflow platforms with AI steps. These let a team connect apps and add a model step, and some now offer agent-style steps in which the model picks among tools. n8n is one example; its documentation describes it as “a fair-code licensed workflow automation tool that combines AI features with business process automation” (n8n docs).
- Automation vendors adding agents. SS&C Blue Prism now describes its offer as “agentic automation” (Blue Prism), and UiPath is described as the platform under PLDT’s systems above.
Before choosing any builder, check what permissions it asks for.
Oversight and limits
An agent can take actions, so the controls matter more than for a chatbot.
Review points. OpenAI’s guide names two triggers that should send a case to a person: exceeding failure thresholds, such as a retry limit, and high-risk actions that are “sensitive, irreversible, or have high stakes.” Refunds, deletions, payments and messages to many customers belong in that group.
Permissions. Give the agent the narrowest access that lets it do the task. An agent that only needs to read order status should not be able to edit prices or send bulk messages.
Logging. Keep a record of what the agent decided, which tools it called and what they returned. Without a log, nobody can explain a wrong action afterwards.
Security. NIST’s Center for AI Standards and Innovation opened a request for information in January 2026 on the security of AI agent systems. It focuses on risks that arise when model outputs are combined with software functions, such as indirect prompt injection, where hidden instructions in a web page or document steer the agent (NIST, 12 January 2026). In practice this means an agent that reads outside content should not also hold broad access to money or customer data.
Personal data. Messages and documents the agent handles often contain personal data. The Data Privacy Act of 2012 applies, and the National Privacy Commission issued Advisory No. 2024-04 on AI systems that process personal data on 19 December 2024. The Philippine data report lists it with other policy items and shows the latest firm-level AI adoption figure found, 3.02 percent of establishments in 2021.
Frequently asked questions
What is the difference between an AI agent and a chatbot?
A chatbot answers messages in a conversation. An AI agent pursues a goal and chooses its own steps and tools, which may include replying in a chat. OpenAI’s guide says simple chatbots that use a model without letting it control the workflow are not agents.
What are the types of AI agents?
The textbook types are simple reflex, model-based reflex, goal-based, utility-based and learning agents. In business tools, agents are more often grouped by task: customer conversations, research and summaries, and operations inside business software.
Can a business build an AI agent without coding?
Often yes. Workflow platforms such as n8n let a team add a model step and connect tools without writing every line. The harder work is writing clear instructions, limiting permissions and checking results.
How much supervision does an AI agent need?
It depends on what the agent can do. Agents that only read and draft need review of their output. Agents that spend money, delete data or contact customers need approval steps, retry limits and logs.
Sources
All sources were read on 12 Oct 2026.
- Anthropic, Building effective agents, 19 December 2024.
- OpenAI, A practical guide to building agents, undated PDF.
- National Institute of Standards and Technology, CAISI issues request for information about securing AI agent systems, 12 January 2026.
- Russell and Norvig, Artificial Intelligence: A Modern Approach, 4th US edition, book website.
- TechNode Global, PLDT reports faster operations from three UiPath AI systems, 10 September 2026.
- n8n documentation, Welcome to the n8n docs.
- SS&C Blue Prism, company site.
- National Privacy Commission, NPC Advisory No. 2024-04, 19 December 2024.