For the past two years, most people’s experience of AI has been conversational: you type something, the AI responds. One turn. One answer. It’s powerful, but it’s fundamentally reactive. Something more significant is now happening: AI systems that can take sequences of actions over time, use tools, browse the web, write and run code, and complete complex multi-step tasks without needing a human to supervise every step.
What Is an AI Agent?
The simplest way to understand an AI agent is to contrast it with a regular AI chatbot. When you ask ChatGPT “How do I create a pivot table in Excel?”, it gives you an answer. That’s it — one question, one response. An AI agent, by contrast, might receive a goal like “Analyse the sales data in this spreadsheet, identify the three lowest-performing product categories, find the most recent customer feedback related to those categories from our support ticketing system, and draft a one-page summary with recommendations” — and then go and do all of that autonomously.
The key characteristics of AI agents are:
- Goal-directed behaviour — they work toward an objective, not just respond to a prompt
- Tool use — they can use external tools: web browsers, code interpreters, databases, APIs, email, calendars
- Multi-step planning — they break complex tasks into steps and execute them in sequence
- Memory and context — they can maintain context across a long task, remembering what they’ve already done
- Autonomy — they complete tasks without needing constant human input
What’s Actually Available Right Now?
Operator (OpenAI)
OpenAI’s Operator is an AI agent that can control a web browser — navigating websites, filling in forms, making purchases, booking appointments and extracting information from web pages. It was launched in early 2025 and is available to ChatGPT Pro subscribers.
Claude Computer Use (Anthropic)
Anthropic’s Computer Use capability allows Claude to control a computer’s graphical interface — not just browse the web but use any desktop application. It can click buttons, type text, navigate menus and use software tools the same way a human would.
Devin and Similar Coding Agents
Cognition’s Devin made headlines as “the first AI software engineer.” It can take a task description, plan and write code, test it, debug errors and deploy solutions — all autonomously.
Multi-Agent Systems
Some of the most interesting developments involve not one AI agent but networks of them, where different agents specialise in different tasks and coordinate with each other.
“Agents are where AI goes from being a very smart calculator to being something that behaves more like a colleague.” — Dario Amodei, Anthropic CEO
The Business Impact: Real and Coming Quickly
Knowledge Work Automation
Tasks that have historically required skilled knowledge workers — research synthesis, report generation, data analysis, contract review, competitive intelligence — are being automated at a speed that is catching many companies off guard.
Software Development
Agentic coding tools are not replacing experienced developers — they are changing what experienced developers spend their time on. Companies report that the ratio of code written by AI vs humans is already above 30% at some organisations.
Customer Operations
AI agents are being deployed in customer-facing roles that go well beyond simple FAQ chatbots. They can look up account information, process refunds, update subscriptions, escalate to human agents when needed.
The Legitimate Concerns
- Errors compound — in a multi-step agentic task, an early mistake can propagate through subsequent steps.
- Security vulnerabilities — agents that can take actions in the world are potentially exploitable by malicious actors.
- Accountability gaps — when an AI agent makes a decision that causes harm, who is responsible?
- Over-reliance — there’s a real risk that humans defer too much to AI agents in high-stakes domains.
How Should Workers and Businesses Respond?
- Experiment now — run real tasks through agent tools and see where they save time and where they fail.
- Define oversight policies — decide which categories of tasks agents can complete autonomously.
- Focus on uniquely human contributions — judgment calls, relationship management, ethical reasoning, creative direction.
- Take security seriously — any agent that can send emails or access sensitive systems needs proper access controls.
Agentic AI is not a future technology. It is a present one, being deployed in production by organisations of all sizes right now.