What is an AI agent?

What is an AI agent?

You are probably familiar with ChatGPT’s empty box.

When it launched in November 2022, it offered a different and better experience than searching on Google. You could explain what you meant, ask for clarification and keep an ongoing conversation. It brought a great deal of knowledge into that exchange, but its answers depended on what it had already been taught. Recent events were out of reach. When it did not know an answer, it could still produce one that sounded convincing. These plausible but false answers are called hallucinations. Without web access, it had no way to check an answer against current online sources. The fix seemed simple: why not let it look things up on the internet and bring the power of search into this new way of asking questions? In March 2023, less than four months after launch, OpenAI introduced tools that included web browsing, allowing it to consult fresh information. It could still make mistakes, but it now had a way to check.

Repeated use revealed another limitation. Imagine adding, “I live in Italy,” to get a more relevant answer. The next day, a new chat begins with advice for someone in the US. It breaks the illusion: “I told you that yesterday.” OpenAI began testing memory in February 2024, allowing useful details to carry between conversations. Giving ChatGPT relevant background could make its answers more useful, but you usually had to gather and supply it yourself. Much of the background an assistant needed, though, was already stored digitally, in emails, appointments and documents. By June 2025, ChatGPT’s research tools could connect to Gmail, Outlook and Google Calendar. With permission, it could consult that information instead of asking you to explain everything. This was another important shift: the assistant could start finding the context it needed about your work, rather than relying on you to bring it every detail.

You could ask it to check when you were free for a meeting. Arranging that meeting was the clear next step, made possible by tools that could act as well as read. Software development showed what this could look like: agents such as Codex could edit code and run checks as part of a task. The approach extended to knowledge work through ChatGPT agent and Claude Cowork. These developments overlapped, with products taking slightly different routes. The goal was to create an AI agent that could use the digital information and tools available to it to carry out a task you describe.

What can it do inside a business?

The potential for a useful personal assistant was becoming clear. The same principles could apply inside a business, with different tools and a different body of knowledge. Those tools might be your email, shared documents, accounting software or project system. An agent could gather updates for a report, prepare a meeting brief or find the information needed for a follow-up. In a July 2026 study of Cowork sessions from May, Anthropic found that business operations and content preparation together accounted for roughly half the sample. People were assembling information, reconciling spreadsheets and preparing presentations. Unlike coding agents used to build software itself, much of this work supported people in applying their expertise. The assistant prepares the picture; your experience helps you decide what to do with it.

Play demo

Hey, could you prepare some notes for my next meeting with Alex?

Reading recent emails

1. Reading recent emails
2. Checking your calendar
3. Preparing your meeting notes

Your notes are ready. I’ve brought together the latest emails, project updates and the points to cover.

Meeting with Alex

One thing: Alex emailed this morning to say today’s time no longer works. I checked your calendar and found two free slots tomorrow.

Which time should I use?

Tomorrow, 10:00–10:30

Tomorrow, 14:30–15:00

Neither works

Connecting it to your software can be the difficult part. Some tools offer suitable connections; others require a software engineer or coordination with the provider. This is part of the work techwaves takes on through its AI Sprint. We help you decide where an agent could be useful, prepare the connections and turn your routines, examples and working rules into guidance it can use. We also help you learn to work with it. That gives the assistant a starting point; the work you do together can make it more useful as it retains relevant preferences and learns from your corrections.

Our proposed experience

The industry is also exploring assistants that become more personal over time. Products such as OpenClaw and Dots reflect that direction: an assistant that keeps working with you across conversations and different forms of communication. At techwaves, we are building toward an experience that fits naturally into your day. You call your assistant, message it in Teams and continue beside an email, with the relevant background carried along. The aim is a familiar helper with a broad view of your business, able to bring more detail into view than you could keep in mind at once.

Imagine reading a customer’s request for a meeting. Your assistant offers two available times and prepares a reply. You adjust the wording, choose a time and approve it. It also knows you like to prepare the day before. When that day comes, it gathers the relevant background, spots a free half-hour in your calendar and sends you a message: “You have time at three. Shall we go through tomorrow’s meeting?” As you work together, your corrections can help it reflect your preferences and recognize familiar routines. It could prepare a recurring briefing, bring an unfinished task back to your attention or review your inbox for a message you may have overlooked. Gathering the background takes less of your day, leaving more room for the judgment, relationships and taste your experience has given you.