The clearest AI agent examples you can use today are coding agents such as Claude Code, OpenAI Codex and GitHub Copilot cloud agent, browser and computer use agents such as Claude in Chrome, customer service agents such as Intercom Fin and Salesforce Agentforce, and research agents such as Gemini Deep Research and Claude's Research feature. Inside a business, the same pattern appears wherever a model is given a goal, a few tools and a loop: approving refunds, working through vendor security reviews, processing insurance claims. In every serious deployment, a person or a hard rule still approves the actions that cannot be undone.
What makes something an AI agent?
OpenAI defines agents as systems that independently accomplish tasks on your behalf and explicitly excludes applications that use a model without letting it control the workflow, naming simple chatbots, single-turn LLMs, or sentiment classifiers. Anthropic's engineering team puts it more mechanically: agents are typically just LLMs using tools based on environmental feedback in a loop.
That gives a simple test for every example in this article. Does the system decide its next step based on what it just found, and can it use tools to change something? If both answers are yes, it is an agent. If it only replies, it is an assistant or a chatbot, which is a different purchase with different risks; we compare the two in AI agent vs chatbot.
Which AI agents can you use today?
Product names in this space change quickly, so every entry below was checked against the vendor's own page in October 2026. GitHub's agent, for example, now appears in its documentation as Copilot cloud agent, and OpenAI's deep research API models o3-deep-research and o4-mini-deep-research are listed with a shutdown date of July 23, 2026.
| # | Agent | Type | What the vendor says it does | Where control sits |
|---|---|---|---|---|
| 1 | Claude Code | Coding | Reads your codebase, edits files, runs commands | A tiered permission system; in manual mode it asks before editing files |
| 2 | OpenAI Codex CLI | Coding | A coding agent from OpenAI that runs locally on your computer | Runs on your machine; a separate cloud agent, Codex Web, runs remotely |
| 3 | GitHub Copilot cloud agent | Coding | Researches a repository, plans changes, and implements them in the background | One branch and one pull request per task, sessions capped at 59 minutes |
| 4 | Claude in Chrome | Browser | Read, click, and navigate websites alongside you | Anthropic says it has safety classifiers but is still risky |
| 5 | OpenAI computer use | Computer use (API) | Operate a browser or desktop interface to fill out forms or test user flows | Your application owns the session, permissions and cleanup |
| 6 | Anthropic computer use tool | Computer use (API) | Screenshot capabilities and mouse/keyboard control | Runs in an environment you provide |
| 7 | Intercom Fin | Customer service | Takes action to update external systems and transfers to agents | Hands off to a human inbox; priced per outcome |
| 8 | Salesforce Agentforce | Customer service | 24/7 autonomous support at enterprise scale | Configured per role with the business knowledge you supply |
| 9 | Gemini Deep Research | Research | Browse up to hundreds of websites and write multi-page reports | Reads your Workspace content only if you choose it |
| 10 | Claude Research | Research | Search across the web, Google Workspace, and any integrations | Read only; results come back with citations |
Coding agents lead because their work is easy to verify. Anthropic's explanation is that code solutions are verifiable through automated tests and agents can iterate on test results, though it adds that human review remains crucial. GitHub's design builds that review in: its cloud agent can only work on one branch at a time and can open one pull request per task, so nothing reaches your main code until someone merges it.
Browser and computer use agents are the most general and the most exposed. They read whatever is on the screen, including text written to manipulate them, which is why Anthropic's own help page says Claude in Chrome is still risky and OpenAI's guide leaves session and permission control to your application.
Customer service agents are where pricing has moved to outcomes. Anthropic noted that several companies charge only for successful resolutions. Intercom's Fin is one: as of October 2026 it is listed at $0.99 per outcome with a 50 outcome monthly minimum when used with another helpdesk. Intercom also states that Fin averages a 76% resolution rate across 12,000+ customers; that is the vendor's own figure.
Research agents show what multiple agents buy you. Anthropic reported that its multi-agent setup, with Claude Opus 4 leading Claude Sonnet 4 subagents, outperformed a single Claude Opus 4 agent by 90.2% on its internal research evaluation, at a much higher token cost. The pattern behind these products is what people mean by agentic workflows: plan, use tools, check, revise.
What are real world examples of AI agents in business?
A case with detailed published figures is Klarna's. In a press release dated February 27, 2024, the company said its OpenAI powered assistant had, in its first month, held 2.3 million conversations, handled two thirds of its customer service chats and done the equivalent work of 700 full-time agents. Klarna also reported a 25% drop in repeat inquiries, resolution times under 2 minutes against 11 minutes before, and an estimated $40 million profit improvement for 2024. Read those as the company's own first month claims, with the profit figure an estimate, not as a benchmark you should expect.
Most business agents are less public and more specific. OpenAI's guide gives three use cases that suit agents, each matching one of its criteria: refund approval in customer service, vendor security reviews, and processing a home insurance claim. The guide also sorts an agent's tools into data tools, such as querying a CRM or reading PDF documents, and action tools, such as sending emails or updating a CRM record. Here is how those three use cases break down, plus one common back office case:
| Use case | Why it suits an agent | Data tools | Action tools | Where a person approves |
|---|---|---|---|---|
| Refund approval | Complex, context-sensitive decisions | Order history, refund policy, customer record | Issue refund, update ticket, reply to customer | Any refund above a set amount |
| Vendor security review | Rulesets too large to maintain by hand | Vendor questionnaire, your security policy, past reviews | Draft findings, request missing evidence | The final pass or fail |
| Home insurance claim | Heavy reliance on unstructured documents | Claim form, photos, policy wording | Request missing documents, set claim status | Every payment |
| Supplier email triage | Messy inbound text, many exception types | Inbox, ERP orders, delivery schedule | Draft reply, tag and route the email | Replies that commit to dates or prices |
The approval column follows OpenAI's own list of high-risk actions that should trigger human oversight: canceling user orders, authorizing large refunds, or making payments. The agent does the reading and the drafting; a person signs off on the money.
What are the 5 types of AI agents?
This question comes from the classic AI textbook, not from today's products. Russell and Norvig's Artificial Intelligence: A Modern Approach describes simple reflex agents, model-based reflex agents, goal-based agents, utility-based agents and learning agents. A thermostat is a simple reflex agent; a route planner that weighs travel time against cost is utility-based.
Every product in the table above is a goal-based agent built on a language model: you give it an objective and it works out the steps. The textbook categories are useful vocabulary, but they will not help you choose between two vendors. The questions that matter for a purchase are which tools the agent gets, what it can do without asking, and how you will check its work.
How do you pick your first agent use case?
Start from the examples that share three traits: the work involves judgment on messy input, the result can be checked, and the expensive mistakes can be put behind an approval step. Anthropic says agents add the most value for tasks with clear success criteria, feedback loops and meaningful human oversight, and OpenAI recommends you validate that your use case can meet these criteria clearly before committing, because otherwise a deterministic solution may suffice.
Then write down, for that one job, the data tools, the action tools and the approval point, exactly as in the table above. If the table has more than a handful of action tools, the scope is too wide for a first project. Our guide on how to build an AI agent walks through the rest of the plan, from the test set to the pilot. If you would like someone to take that plan through to production, Qubba offers AI agent development scoped per process, with a fixed written quote after a 30 minute call.
Frequently asked questions
Is ChatGPT an AI agent?
It depends on what it is doing.
Answering a question in a chat window is not agentic by OpenAI's own definition, which excludes simple chatbots and single-turn uses.
OpenAI's current product docs describe a broader role, in which ChatGPT can gather context, take action, and produce something useful.
When it uses tools to carry out a multistep task, it is working as an agent.
What are the top AI agents?
There is no neutral ranking, because agents are built for different jobs.
For software work, look at the coding agents in the table; for long reports, the research agents; for customer chats, the service agents.
Compare them on the tasks you actually have, with your own test cases, rather than on a vendor's headline figure.
Are AI agents safe to use with company data?
They can be, if access is limited to what the task needs.
The products above all restrict something: Claude Code asks before edits in manual mode, GitHub's agent works through pull requests, and Gemini Deep Research reads Workspace content only if you allow it.
Apply the same thinking to anything you build: minimum tools, minimum permissions, and a person in front of irreversible actions.
Do AI agents replace employees?
Some companies report that agents take over a large share of routine volume, as Klarna did for its customer service chats.
In most deployments the agent handles the reading, drafting and routine cases, while people handle exceptions and approve the actions that carry financial or legal weight.