For several years, most people experienced generative AI through a chat box. You typed a question, the AI produced an answer, and nothing happened unless you took the next step yourself.

Then, seemingly all at once, everything became an “AI agent.” Software companies announced teams of agents for sales, marketing, customer service, and operations. Social media filled with promises that anyone could deploy a collection of digital workers to run a business. The viral rise of OpenClaw in early 2026 pushed the idea even further into the mainstream.

The trouble is that many of these products do very different things. Some are ordinary automations with an AI step added. Some are assistants that can search, analyze, and draft. Others can make decisions, operate software, and carry work forward with limited supervision.

All of them may be marketed as agents. That makes a simple question surprisingly difficult to answer: What actually is an AI agent?

AI escaping the chat box

One useful way to recognize agentic behavior is to look for AI “escaping the chat box.”

A chatbot can help a customer service employee think through a complaint or draft a response. An agent could take the next steps inside the company’s systems: review the customer’s order history, apply the relevant policy, prepare the response, and potentially send it. Depending on the rules it has been given, it might still ask for approval before issuing a large refund.

That ability to operate within the tools where the work happens is a major step beyond producing an answer in a chat box.

Working across tools, tabs, and apps is a strong sign of agency. The same goes for continuing a task without waiting for approval at every step. There are exceptions. A coding agent can do meaningful work inside one development environment, and many capable agents still pause at approval gates. A useful test is how often the person has to direct the next action.

Why nobody agrees on the definition

Part of the confusion comes from the fact that major AI companies draw the boundary in different places.

Anthropic distinguishes between workflows, where models and tools follow predefined paths, and agents, where the model dynamically directs its own process and tool use. Under this definition, an agent operates in a loop: it plans, acts, observes the result, adjusts, and continues until it completes the task or needs help. For example, it might try to reschedule a service appointment, discover that the preferred technician is unavailable, check another location, and offer the customer a new option.

OpenAI emphasizes independent workflow execution: an agent manages the process, chooses tools, corrects mistakes, and knows when to stop or return control to a person. For example, it could process a refund request from start to finish, escalating only when the amount exceeds company policy.

Google uses a broader definition centered on pursuing a goal through reasoning, planning, memory, action, and some degree of autonomy. It may even describe a collaborative AI assistant as a type of agent. For example, an assistant that remembers a manager’s preferences and coordinates a weekly schedule could qualify even when the manager remains closely involved.

Microsoft often frames agents as systems that perceive their environment, reason about a goal, and take action with tools. In practice, that can include anything from a narrowly scoped service agent to a system coordinating several specialized agents.

Each company draws the line around “AI agent” a little differently, which helps explain why the term can feel so inconsistent. We went through something similar when “AI” first entered everyday conversation. The goalpost for what counted as real AI kept moving until the term became familiar enough that people focused more on what the technology could do.

Automation, agentic workflow, or agent?

A familiar business task makes the distinction easier to see.

Imagine that a customer submits an after-hours service request to a multi-location HVAC company.

Traditional automation: The system creates a work order, selects a branch based on the ZIP code, and sends a standard confirmation. The rules and path were specified in advance.

Agentic workflow: AI interprets the customer’s description, recognizes signs of an emergency, and routes the request into a predefined emergency process. AI is making a judgment, but the surrounding path is still largely designed by people.

AI agent: The system receives the goal “handle and schedule this request.” It checks the customer history, decides whether the issue is urgent, reviews technician availability and location, contacts the customer, books the appointment, and updates the work order. If no technician is available, it changes its approach or asks a dispatcher for help.

In other words, automation follows a path that people designed. An agentic workflow uses AI judgment at certain points in that path. An agent has more freedom to decide how the task should be completed.

Three dimensions of agency

Agency is better understood as a spectrum. These three questions help show where a system sits on it.

1. Decision independence

Can the system decide what to do next, or does it simply follow a fixed sequence? A workflow becomes more agentic when the AI can choose among possible actions, adapt when something changes, and recover from a failed attempt.

2. Environmental reach

Can it only produce an answer, or can it perceive and operate within the systems where work happens? Browser control, computer use, APIs, files, calendars, inboxes, and business software give the AI an environment in which to act.

3. Execution authority

Can it only recommend an action, or is it allowed to carry one out? Drafting an email and sending it are different levels of authority. So are identifying the least expensive supplier, adding an item to a cart, and completing the purchase.

These dimensions do not always rise together. A deep-research tool may make many independent decisions about where to search and how to investigate, yet have little authority outside its research environment. A simple automation may have permission to send thousands of messages but make almost no independent decisions. Neither capability alone tells the whole story.

One purchase across the spectrum

Consider a restaurant group that needs to replace several point-of-sale tablets.

An AI assistant might summarize the requirements, research compatible models, and create a comparison table. That is useful, but the work still ends with information in a chat.

A more agentic system could open supplier websites, check current prices and availability, compare delivery dates for each restaurant, choose the best eligible option, and add the correct quantities to a cart. It has crossed from advising into operating.

A business could still require a manager to approve the payment. The agent would have less authority, while the research, comparison, and cart-building steps would remain agentic.

Is it really agentic?

The current wave of agent marketing makes the terminology even harder to trust. A company can label each automated step as a separate agent and claim to offer an “army” of digital employees. A social media post can describe a carefully supervised set of templates and integrations as agents that run an entire business.

The label is less useful than a few practical questions:

• What goal can the system accept?
• Which decisions can it make on its own?
• Which tools and information can it access?
• What actions can it actually execute?
• Can it observe results and adjust its approach?
• When does it need human approval?
• What happens when it is uncertain or something goes wrong?

If a vendor cannot answer those questions clearly, “agent” may be doing more work in the marketing than in the product.

“Agentic” may be more useful than “agent”

There may never be one definition that cleanly separates agents from everything else. The products are changing too quickly, the capabilities exist on a spectrum, and the tools people already know have gradually added research, tool use, browser control, computer use, and other agentic features. The transition did not happen at a single obvious moment.

That is why “agentic” is often the more useful word. It shifts the conversation away from a product category and toward observable behavior.

An agentic system can decide how to pursue a goal, act within an environment, evaluate what happened, and continue without constant direction. The degree of agency depends on how independently it can make decisions, how far it can reach into the tools where work happens, and how much authority it has to act.

When someone claims that an AI agent can run a workflow, or even an entire business, ask for specifics. What can it see? Which decisions can it make? What actions can it take? Where does a person still step in?