AI agentsGuide

What are AI agents

Understand what makes an AI system an agent, the roles of models, tools and context, and how each action guides the next step.

An AI agent is a system that uses artificial intelligence to choose and carry out actions toward a goal, taking the results of those actions into account.

Imagine asking an AI to prepare for a meeting. It can explain how to put together a useful agenda. It can also summarize the client history you provide. An agent can go further: use the tools available to look up that history, spot an unresolved issue, investigate it and adjust the agenda based on what it finds.

The central idea is the ability to guide part of the work between a request and a finished result. The system receives a goal and chooses its next steps as it works, within the access and limits it has been given.

To understand what that means, we need to distinguish three things that often appear together: talking to an AI, automating a sequence and delegating a task to an agent.

What makes an AI system an agent?

Consider this request: “Prepare an agenda covering the topics we need to discuss with this client.”

WHAT HAPPENS AFTER THE REQUEST

Who decides what happens next?

  1. Chat with an AI without tools

    You decide what to look up.

    The AI helps interpret the information and draft a response.

  2. Automation with a predefined workflow

    The sequence is configured in advance.

    The system follows the steps and branches according to preset rules.

  3. AI agent

    Part of the path is chosen during the task.

    The system decides what to look up or do based on the goal and what it finds.

Comparison by Agora. The difference is when and by whom the next steps are decided.

The distinction is between following a defined sequence and being able to choose part of that sequence while the task is underway.

Anthropic uses this distinction between predefined workflows and agents in its guide Building effective agents. Terminology varies across the industry; here, the distinction helps explain how the system behaves.

These approaches can also work together. A fixed routine can start a task, an agent can investigate an issue and another routine can send the result for review.

Chat is the interface you use to communicate. An agent can operate within it. To recognize one, look at the work it can direct: which actions it chooses, which results it observes and how it uses that information to continue.

The AI model is one part of the agent

A language model interprets instructions and produces responses. To look up a document, create a file or update a record, it needs to be connected to a system that provides those capabilities.

In agents built around language models, the model helps interpret the task and select actions. The surrounding software runs the tool calls, returns their results and applies the controls.

INSIDE THE SYSTEM

The parts of an AI agent

Goal and instructionsPrepare the agenda for the client meeting.
AI AGENTAgent

Context

History, messages and lookup results.

informs

AI model

Interprets the situation and selects the next lookup or action.

requestsreturns

Tools

Software looks up records and drafts the agenda.

The tool's result returns to context and guides the next decision.

Controls throughout the task Looking up records and preparing a draft are allowed. Sending it to the client requires approval.
DeliverableAn agenda for review.
Simplified diagram by Agora. The model is part of the agent; the surrounding system connects tools and applies controls.

A tool might allow the agent to read documents, check a calendar or create an agenda. When the agent selects a tool, it makes a request that the software executes. The result comes back as information for the next step.

This explains why two agents using the same model can produce different work. One might have access to up-to-date history; another, only an old summary. One might be able to prepare a document; another can also share it. The model provides part of the capability, and the system determines how it can be used.

What role does context play?

Context is the information the model receives to work with at a particular moment: instructions, messages, selected documents and tool results, for example.

When preparing for a meeting, an old note might say the client is satisfied. A recent support ticket might reveal an unresolved problem. Access to that second record changes what deserves attention.

Connecting an agent to a document library does not mean it considers every document in every decision. Information must be retrieved and selected for the model. Selecting and maintaining that information is part of context engineering, as described by Anthropic.

Context is also different from persistent memory. Information used now can only be retrieved on another occasion if the system has a way to store it and bring it back. Continuity across tasks depends on that design.

How does the agent choose its next step?

One important pattern is the action and observation loop: choose an action, receive its result and use that information to decide how to continue.

The ReAct research, by researchers at Princeton and Google, explored combining reasoning with actions that interact with external sources. This loop helps explain how new information can change the direction of a task.

THE ACTION-AND-OBSERVATION LOOP IN A FICTIONAL EXAMPLE

New information changes the next step

  1. Person

    Prepare an agenda with the topics we need to discuss with this client.

  2. Review the account historyAgent
    Client history

    There is interest in expanding the contract.

    Next step: Check recent records before drafting the agenda.

  3. Check recent support ticketsAgent
    Support ticket

    A product issue was reported.

    Next step: Find out whether the issue has been resolved.

  4. Check the ticket updateAgent
    Support update

    The issue is still open.

    Agenda change: Address its impact before discussing an expansion.

  5. Incorporate a human reviewPerson
    Account owner

    “We agreed on a temporary workaround over the phone.”

    Agenda change: Adjust the agenda to check whether the workaround worked.

Fictional example by Agora illustrating the action-and-observation loop. Each result can prompt a new lookup or change the deliverable.

The example shows the loop in action: one lookup produces information, that information guides another lookup and human input changes the deliverable. Part of the path was determined during the task.

An action can also fail. If a lookup returns no data, the system needs to distinguish missing information from confirmation. If a tool reports that it created a document, the execution result needs to support that conclusion.

The same pattern can appear in different tasks. In these fictional examples, notice how new information changes the next action:

TWO TASKS, THE SAME PATTERN

The result guides the next action

Why hasn't my order arrived?

Customer supportAgent
  1. LookupShipment tracking
  2. FindingThe address is incomplete.
  3. Next stepCheck the customer record and ask for confirmation before suggesting a correction.

Compare these two software options.

ResearchAgent
  1. LookupProduct documentation
  2. FindingOne feature requires a separate plan.
  3. Next stepVerify that requirement before drawing a conclusion.
Fictional examples. An agent can adjust part of its path based on what it finds. If every step and branch is fixed in advance, the same task can be a predefined workflow.

Does an agent have to work alone?

The degree of autonomy is part of the system's configuration and can vary within the same task.

An agent can choose which records to investigate and prepare an agenda, while sending it to the client requires approval. Asking for help, receiving a correction or referring a decision to a person are ways to keep the work moving within those limits.

There are also conditions for stopping: completing the deliverable, encountering a blocker or reaching a time, cost or attempt limit. Autonomy does not imply endless execution or unrestricted access to tools.

These limits need to exist in the software. Asking the agent in an instruction to avoid changing a record has a different effect from providing it with a read-only tool.

Being an agent describes a way of carrying out tasks. Quality depends on how the system interprets the goal, selects information, uses tools and checks what happened.

When preparing for the meeting, it can still confuse an intention with a commitment, use an outdated record or miss an unresolved issue. A well-written response can contain those errors. Sources and action results help people examine the work and correct its course.

When a product is described as having an “AI agent,” the definition becomes more concrete if we can answer: what goal does it receive, which actions can it choose and how does the result of one action guide the next?

The connection between goal, action and observation is the central idea of an AI agent. Understanding it makes it easier to see what the system actually does and where human participation enters the process.

Agora

People and agents. One team.

Agents that work with your team across conversations, meetings, and tools.

Join the waitlist