What are AI agents? A plain-language guide for teams
85% of Fortune 100 companies choose Asana¹
Summary
An agent pulls information from your other systems to finish a job, like gathering project updates and drafting a status report for someone to approve. Its permissions determine which actions it can take.
This 10-minute guide goes over different types of agents and how they work. It also explores workplace examples and how agents support AI work management.
Master the AI Revolution at Work
Learn how leading organizations are turning AI from a tool into a trusted teammate. Access our latest research report to discover five proven secrets of AI-mature organizations and transform your workplace.
Why AI agents matter now
Generative AI helps you produce a first draft. AI agents can take the next actions in a workflow, such as saving that draft to a project and assigning it to a reviewer. You define the steps they handle and the approvals they need before moving work forward.
From generative AI to agentic AI
Generative AI creates content in response to a prompt, such as drafting a reply to a customer request. Agentic AI uses AI to plan and carry out the steps needed to complete a task. For example, an agent can look up the customer’s account history in your CRM, draft a reply using those details, and assign it to the right person for review.
The misconception: AI agents vs chatbots
The difference is what the tool does after you hit enter. A chatbot responds to your message. An AI agent takes actions through connected tools to complete a task, such as checking a customer record and assigning a follow-up. Some chatbots include agent capabilities, so look at what the tool does, not just its chat interface.
Types of AI agents
The five classic types of AI agents range from systems that follow fixed rules to agents that learn from feedback. These categories can overlap. A learning agent might also compare ways to reach a goal by weighing their likely benefits and costs, such as choosing a delivery route based on travel time and fuel use.
Type | How it works | Example |
|---|---|---|
Simple reflex agent | Follows a rule based on the information it receives now | Routes a request marked “billing” to the finance team |
Model-based reflex agent | Uses current information and an internal picture of the situation to choose an action | Checks earlier messages to avoid asking a customer for details they already provided |
Goal-based agent | Chooses steps that move it toward a defined outcome | Identifies gaps in a research brief and searches for information to fill them |
Utility-based agent | Compares possible outcomes using criteria that define the most useful result | Recommends a meeting time by weighing attendance against scheduling delays |
Learning agent | Uses feedback or experience to improve how it performs a task | Gets better at sorting support requests as the team corrects it |
A single agent handles a task with its available tools. A multi-agent system coordinates several agents, often assigning each a different role. One agent might gather research while another checks the findings. Any of the five agent types can work alone or as part of a multi-agent system.
How AI agents work
AI agents use a repeating process to move from a goal to a result. They gather relevant information and then use connected tools to take action. After each action, they check the outcome before continuing.
The actions an agent can take depend on its tools and permissions. An agent with read-only access might prepare a report. An agent with permission to update records could also add that report to a project. You decide where the agent can act independently and when it needs human input.
The agent loop: perceive, reason, act, observe
An agent begins with a goal, such as preparing a weekly project update. It gathers context from the sources it can access and uses that information to plan its next step. Then it takes an action through a tool and checks the result. It repeats this loop until it meets the goal, reaches a limit, or needs help.
Each stage serves a different purpose:
Stage | What the agent does | Project-update example |
|---|---|---|
Perceive | Gathers information relevant to the goal. | It reads recent task updates and compares them with the project timeline. |
Reason/plan | Uses the available context to choose its next step. | It identifies a missed milestone and decides to check the latest comments for an explanation. |
Act | Uses a connected tool within its permissions. | It retrieves comments on the tasks tied to the missed milestone and drafts a project update explaining the delay. |
Observe | Checks the outcome before continuing. | It checks whether a draft covers the requested reporting period and flags any missing information. |
The checking step matters because taking an action doesn’t guarantee success. A search may return an outdated document, or an app may reject an update. An agent that assumes every action worked will build the rest of the task on a bad step.
What an AI agent is made of: model, tools, memory, and a goal
An AI agent combines a model with tools that let it interact with its working environment. Memory helps it retain relevant context, while a goal directs its decisions. A person sets guardrails around the process so the agent understands which actions it can take and when it should stop.
These parts work together, but each handles a different job.
Part | Role in the agent |
|---|---|
Model | Interprets the request and helps decide which step could advance the goal |
Tools | Let the agent retrieve information or take action, such as searching documents or updating a task |
Memory | Holds context from the task, including earlier results. Some systems retain context across sessions; others limit it to the current run. |
Goal and guardrails | Define the intended outcome and the boundaries for pursuing it, including access restrictions and approval requirements |
AI agent examples
An AI agent can turn a request into action, such as checking an order’s progress and assigning a follow-up to the right person. The tasks it handles depend on the tools you connect and the access you give it, from daily tasks to complex asks.
Research agent: A team lead asks for a competitive brief. The agent searches approved sources and compares the findings with the brief’s requirements. If it finds a gap, it searches again before drafting the document. It includes source links so the team can check the evidence.
Support agent: An agent reviews an incoming request and checks the customer’s account history. It categorizes the issue and drafts a reply using approved guidance. When a request falls outside that guidance, it hands the case to a person.
Marketing-ops agent: An agent gathers campaign updates from connected project records and prepares a status report. It flags conflicting information for review instead of guessing which update to trust. A campaign manager checks the report before the agent shares it.
How teams use AI agents at work
In Asana, you assign work to AI Teammates much like you would to a colleague. They draw on the Work Graph®, which connects tasks and projects to the people responsible for them and the goals they support. That gives an AI Teammate context about who owns the work and how it fits into a larger plan.
For example, ask an AI Teammate to turn a launch brief into a project plan with tasks and owners, or consolidate feedback on campaign visuals into a revision list. The work stays in your shared projects, where you can review progress and guide the next steps. Explore AI workflow automation to build these actions into your existing processes.
AI agents at work: what teams are seeing
At performance marketing agency Amsive, handing a new client from sales to onboarding meant chasing down information. Account details lived in Salesforce, while employees had to search messages and documents for the rest. Details often went missing along the way.
Now, when a deal closes in Salesforce, an integration creates a client task in Asana. An AI Teammate gathers the client’s information from connected tools and puts it into a profile for the Operations Manager to review. Once the manager launches the onboarding project, the AI Teammate creates tasks based on the services the client purchased.
“The result is institutional knowledge that scales, that doesn't live in just people's heads,” said Adam Lozada, Digital Workplace and AI Enablement Lead at Amsive.
Amsive estimates the workflow saves more than 3,800 hours a year by cutting time spent searching for information and filling out duplicate onboarding forms. The AI Teammate handles the gathering and setup, while the Operations Manager checks the details before onboarding begins.
Read the Amsive storyPut AI agents to work with Asana
Spend less time gathering information before the work begins. AI Teammates in Asana use the details in your projects to help turn a launch brief into assigned tasks or pull scattered feedback into a revision list. See what you could hand off while keeping your team in control of the decisions that need a human.
Start free trialFrequently asked questions about AI agents
ChatGPT can function as an AI agent, depending on how you use it. In a standard conversation, it responds to prompts as a generative AI assistant. In agent mode, it uses tools to take actions toward a goal, such as navigating websites or editing spreadsheets. The distinction lies in completing tasks rather than just generating answers.
The five classic types of AI agents are simple reflex, model-based reflex, goal-based, utility-based, and learning agents. Simple reflex agents follow rules; model-based reflex agents also track context. Goal-based agents pursue an outcome, while utility-based agents weigh possible results. Learning agents use feedback or experience to improve their performance on a task.
Alexa has some agent-like features, such as acting on voice requests and controlling connected devices. Traditional Alexa mostly waits for a command and carries it out. Alexa+ adds more agentic capabilities, using connected services to complete tasks on your behalf. The distinction depends on the version and the task you ask it to handle.
Generative AI creates content in response to a prompt, while agentic AI plans steps and takes action through tools to pursue a goal. For example, generative AI can draft a project update, while an agent gathers the latest task information and shares the update after approval. Agents often use generative AI as part of that process.
Yes, with clear limits and oversight. Give agents access only to the data they need, and require approval for sensitive changes, such as deleting records or changing someone’s permissions. Your team should also decide who checks the agent’s work and handles mistakes. To learn more about managing access, read about AI governance in Asana.
Choose one repeatable task, such as drafting a project update for your review. Decide what a useful draft should include and give the agent access to the project information it needs. Check the results before adding more responsibilities. For ideas on where to begin, see AI workflow automation in Asana.
