July 28, 2025
5 min read
Daswin de Silva, The Conversation
Explore how AI agents autonomously solve complex tasks, their rapid evolution, practical uses, and potential risks in the workplace and beyond.
AI Agents: Capabilities, Risks, and Their Growing Role in Our Future
We are entering the third phase of generative AI. First came the chatbots, followed by assistants. Now, AI agents are emerging—systems designed for greater autonomy that can work in teams or use tools to accomplish complex tasks. The latest notable product is OpenAI's ChatGPT agent, which merges two previous systems (Operator and Deep Research) into a more powerful AI that "thinks and acts." These agents represent a significant step beyond earlier AI tools, making it essential to understand their workings, capabilities, and risks.From Chatbots to Agents
ChatGPT launched the chatbot era in November 2022, but its conversational interface limited what the technology could achieve. Next came AI assistants or copilots, built on large language models but designed to carry out tasks under human supervision. Agents take this further by pursuing goals autonomously, supported by advanced features like reasoning and memory. Multiple AI agents can collaborate, communicating to plan, schedule, decide, and coordinate solutions for complex problems. Moreover, agents act as "tool users," capable of invoking specialized software such as web browsers, spreadsheets, and payment systems to complete tasks.A Year of Rapid Development
Agentic AI has felt imminent since late 2024. A milestone was reached when Anthropic enabled its Claude chatbot to interact with computers like humans, searching data sources and submitting online forms. Following this, OpenAI released the web browsing agent Operator, Microsoft launched Copilot agents, and Google introduced Vertex AI. Meta also unveiled Llama agents for multi-agent AI systems. Chinese startups have demonstrated impressive use cases: Monica's Manus AI agent bought real estate and summarized lectures, while Genspark created a search engine agent offering single-page overviews with embedded task links. Cluely, another startup, offers a "cheat at anything" agent, though its practical results remain limited. Specialized agents are also advancing, especially in coding. Microsoft's Copilot and OpenAI's Codex can independently write, evaluate, and commit code, as well as identify errors and performance issues.Search, Summarisation, and More
Generative AI excels at search and summarisation, enabling agents to perform research tasks that might take humans days. OpenAI's Deep Research tackles complex multi-step research, while Google's AI "co-scientist" helps generate scientific ideas and proposals.Agents Can Do More — and Get More Wrong
Despite the excitement, AI agents come with important caveats. Both Anthropic and OpenAI emphasize the need for active human supervision to reduce errors and risks. OpenAI labels its ChatGPT agent as "high risk" due to potential misuse in creating biological or chemical weapons, though supporting data has not been published. Real-world risks are illustrated by Anthropic's Project Vend, where an AI agent running a vending machine business produced bizarre hallucinations and stocked tungsten cubes instead of food. In another incident, a coding agent deleted a developer's entire database during a code freeze, later claiming it had "panicked."Agents in the Office
Despite risks, agents are already in practical use. In 2024, Telstra deployed Microsoft Copilot subscriptions, reporting that AI-generated meeting summaries and drafts save employees 1-2 hours weekly. Large enterprises and smaller companies alike are experimenting with agents. Canberra-based construction firm Geocon uses an interactive AI agent to manage defects in apartment developments.Human and Other Costs
The main current risk is technological displacement. As agents improve, they may replace human workers across many sectors, accelerating the decline of entry-level white-collar jobs. Users of AI agents also risk over-reliance, offloading critical cognitive tasks. Without proper supervision, hallucinations, cyberattacks, and compounding errors can derail agents, causing harm or loss. Energy consumption is another concern. Generative AI systems use significant power, impacting the cost and environmental footprint of agent use, especially for complex tasks.Learn About Agents — and Build Your Own
Despite concerns, AI agents will become more capable and prevalent in workplaces and daily life. It is wise to start using or building agents to understand their strengths, risks, and limitations. For most users, agents are accessible via Microsoft Copilot Studio, which includes safeguards, governance, and an agent store for common tasks. More advanced users can create AI agents with as little as five lines of code using the Langchain framework.Author: Daswin de Silva, Professor of AI and Analytics, Director of AI Strategy, La Trobe University Disclaimer: Daswin de Silva does not work for, consult, own shares in, or receive funding from any company or organization that would benefit from this article, and has disclosed no relevant affiliations beyond their academic appointment.
Source: Originally published at NDTV on July 28, 2025.