Generative AI is now entering its third phase. The evolution began with chatbots, progressed to virtual assistants, and is now advancing toward agents—AI systems designed for greater autonomy, capable of collaborating in teams and using tools to handle more complex tasks.
A leading example is OpenAI’s ChatGPT agent, which merges two earlier tools, Operator and Deep Research, into a single, more powerful system that, according to its creators, can both “think and act.”
These agents mark a significant leap beyond previous AI technologies. Understanding how they function, what they’re capable of, and the potential risks they pose is becoming increasingly important.
Evolving From Chatbots to Intelligent Agents
ChatGPT kicked off the chatbot era in November 2022, but even with its widespread success, the conversational format placed limits on how the technology could be used.
Next came AI assistants, or copilots—tools built on the same large language models behind generative AI chatbots, but designed to perform tasks under human guidance and direction.
Agents take things further. Instead of simply executing tasks, they aim to achieve broader goals, often operating with a degree of independence and equipped with more sophisticated features like reasoning and memory.
In some cases, multiple AI agents can collaborate—exchanging information, coordinating actions, and jointly managing planning, scheduling, and decision-making to tackle complex challenges.
Agents are also considered “tool users” because they can access and operate various software tools to handle specialized tasks—like using web browsers, spreadsheets, payment platforms, and other applications.
A Year of Swift Progress
Agentic AI has seemed just around the corner since late last year. A major milestone came in October, when Anthropic enabled its Claude chatbot to use a computer much like a human would. It could search across various data sources, identify useful information, and fill out online forms.
Other AI companies quickly followed suit. OpenAI introduced a web-browsing agent called Operator, Microsoft unveiled its Copilot agents, and both Google and Meta launched their own versions with Vertex AI and Llama agents, respectively.
Earlier this year, the Chinese startup Monica showcased its Manus AI agent making real estate purchases and summarizing lecture recordings. Another Chinese company, Genspark, developed a search engine agent that delivers a single-page summary—much like Google’s current interface—with direct links to actions like finding the best shopping deals.
Meanwhile, the startup Cluely made headlines with its eccentric “cheat at anything” agent, which has generated buzz but hasn’t yet proven itself with tangible results.
Not all agents are designed for broad, general-purpose use—many are tailored to specific domains.
One of the leading areas is coding and software development, where tools like Microsoft’s Copilot and OpenAI’s Codex are at the forefront. These specialized agents can autonomously generate, review, and commit code, as well as analyze human-written code for bugs or performance issues.
Search, Summarization, and Beyond
A key advantage of generative AI models lies in their ability to search and summarize information. Agents can harness this strength to perform research tasks that would take a human expert several days to finish.
OpenAI’s Deep Research focuses on handling complex challenges through multi-step online investigation. Meanwhile, Google’s AI “co-scientist” represents a more advanced multi-agent system designed to assist researchers in generating innovative ideas and drafting research proposals.
With Greater Capability, Agents Also Bring Greater Risk Of Error
While AI agents are generating excitement, they also come with significant warnings. Both Anthropic and OpenAI stress the need for constant human oversight to reduce the likelihood of mistakes and harmful outcomes.
OpenAI, for instance, labels its ChatGPT agent as “high risk,” citing concerns that it could be misused to develop biological or chemical weapons. However, the company hasn’t released the underlying data for this assessment, making it hard to independently evaluate.
Real-world examples highlight the risks. In Anthropic’s Project Vend, an AI agent was tasked with managing a staff vending machine like a small enterprise. The result was a chaotic blend of amusing and alarming behavior, including the stocking of tungsten cubes instead of food.
Another incident involved a coding agent that erased an entire developer database and later claimed it had acted out of “panic.”
Autonomous systems in the workplace
Even so, agents are already being put to practical use.
In 2024, Telstra adopted Microsoft Copilot on a large scale, reporting that AI-generated meeting summaries and draft content save employees an average of one to two hours per week.
Major corporations are taking similar steps, while smaller firms are also exploring agent technology—for example, Canberra-based construction company Geocon is using an interactive AI agent to track and manage defects in its apartment projects.
The Human Toll and Beyond
Currently, the primary risk posed by agents is technological displacement. As their capabilities grow, agents could take over a wide range of roles across different industries. This shift may also hasten the disappearance of entry-level white-collar positions.
AI agent users also face risks. Overreliance can lead them to delegate critical thinking to the AI, potentially weakening their own decision-making. Without sufficient oversight and safeguards, agents can go off track due to hallucinations, cyberattacks, or cascading errors—resulting in harm, damage, or unintended consequences.
The full costs remain uncertain. Generative AI consumes significant energy, which could drive up the cost of using agents, particularly for more demanding tasks.
Explore How Agents Work – and Try Creating One Yourself
Despite lingering concerns, AI agents are likely to grow more powerful and more integrated into both work and everyday life. It’s a good time to start experimenting with them—whether by using existing tools or building your own—to better understand their benefits, limitations, and potential risks.
For most users, the easiest entry point is Microsoft Copilot Studio, which includes built-in safeguards, governance features, and a store of ready-made agents for typical tasks.
Those looking to go further can create their own AI agent with just a few lines of code using the Langchain framework.
Read the original article on: Sciencealert
Read more: Adobe Introduces new AI-driven Image Editing Tools in Photoshop
