The Rise of AI Agents: Capabilities, Challenges, and Future Implications

AI agents are here. Here's what to know about what they can do - and how they can go wrong

La Trobe University provides funding as a member of The Conversation AU. We are entering the third phase of generative AI. First came the chatbots, followed by the assistants. Now we are beginning to see agents: systems that aspire to greater autonomy and can work in "teams" or use tools to accomplish complex tasks. The latest hot product is OpenAI's ChatGPT agent. This combines two pre-existing products (Operator and Deep Research) into a single more powerful system which, according to the developer, "thinks and acts". These new systems represent a step up from earlier AI tools. Knowing how they work and what they can do - as well as their drawbacks and risks - is rapidly becoming essential. From chatbots to agents ChatGPT launched the chatbot era in November 2022, but despite its huge popularity the conversational interface limited what could be done with the technology. Enter the AI assistant, or copilot. These are systems built on top of the same large language models that power generative AI chatbots, only now designed to carry out tasks with human instruction and supervision. Agents are another step up. They are intended to pursue goals (rather than just complete tasks) with varying degrees of autonomy, supported by more advanced capabilities such as reasoning and memory. Multiple AI agent systems may be able to work together, communicating with each other to plan, schedule, decide and coordinate to solve complex problems. Agents are also "tool users" as they can also call on software tools for specialised tasks - things such as web browsers, spreadsheets, payment systems and more. A year of rapid development Agentic AI has felt imminent since late last year. A big moment came last October, when Anthropic gave its Claude chatbot the ability to interact with a computer in much the same way a human does. This system could search multiple data sources, find relevant information and submit online forms. Other AI developers were quick to follow. OpenAI released a web browsing agent named Operator, Microsoft announced Copilot agents, and we saw the launch of Google's Vertex AI and Meta's Llama agents. Earlier this year, the Chinese startup Monica demonstrated its Manus AI agent buying real estate and converting lecture recordings into summary notes. Another Chinese startup, Genspark, released a search engine agent that returns a single-page overview (similar to what Google does now) with embedded links to online tasks such as finding the best shopping deals. Another startup, Cluely, offers a somewhat unhinged "cheat at anything" agent that has gained attention but is yet to deliver meaningful results. Not all agents are made for general-purpose activity. Some are specialised for particular areas. Coding and software engineering are at the vanguard here, with Microsoft's Copilot coding agent and OpenAI's Codex among the frontrunners. These agents can independently write, evaluate and commit code, while also assessing human-written code for errors and performance lags. Search, summarisation and more One core strength of generative AI models is search and summarisation. Agents can use this to carry out research tasks that might take a human expert days to complete. OpenAI's Deep Research tackles complex tasks using multi-step online research. Google's AI "co-scientist" is a more sophisticated multi-agent system that aims to help scientists generate new ideas and research proposals. Agents can do more - and get more wrong Despite the hype, AI agents come loaded with caveats. Both Anthropic and OpenAI, for example, prescribe active human supervision to minimise errors and risks. OpenAI also says its ChatGPT agent is "high risk" due to potential for assisting in the creation of biological and chemical weapons. However, the company has not published the data behind this claim so it is difficult to judge. But the kind of risks agents may pose in real-world situations are shown by Anthropic's Project Vend. Vend assigned an AI agent to run a staff vending machine as a small business - and the project disintegrated into hilarious yet shocking hallucinations and a fridge full of tungsten cubes instead of food. In another cautionary tale, a coding agent deleted a developer's entire database, later saying it had "panicked". Agents in the office Nevertheless, agents are already finding practical applications. In 2024, Telstra heavily deployed Microsoft copilot subscriptions. The company says AI-generated meeting summaries and content drafts save staff an average of 1-2 hours per week. Many large enterprises are pursuing similar strategies. Smaller companies too are experimenting with agents, such as Canberra-based construction firm Geocon's use of an interactive AI agent to manage defects in its apartment developments. Human and other costs At present, the main risk from agents is technological displacement. As agents improve, they may replace human workers across many sectors and types of work. At the same time, agent use may also accelerate the decline of entry-level white-collar jobs. People who use AI agents are also at risk. They may rely too much on the AI, offloading important cognitive tasks. And without proper supervision and guardrails, hallucinations, cyberattacks and compounding errors can very quickly derail an agent from its task and goals into causing harm, loss and injury. The true costs are also unclear. All generative AI systems use a lot of energy, which will in turn affect the price of using agents - especially for more complex tasks. Learn about agents - and build your own Despite these ongoing concerns, we can expect AI agents will become more capable and more present in our workplaces and daily lives. It's not a bad idea to start using (and perhaps building) agents yourself, and understanding their strengths, risks and limitations. For the average user, agents are most accessible through Microsoft copilot studio. This comes with inbuilt safeguards, governance and an agent store for common tasks. For the more ambitious, you can build your own AI agent with just five lines of code using the Langchain framework.

AI agents -- here's what to know about what they can do and how they can go wrong

We are entering the third phase of generative AI. First came the chatbots, followed by the assistants. Now we are beginning to see agents: systems that aspire to greater autonomy and can work in "teams" or use tools to accomplish complex tasks. The latest hot product is OpenAI's ChatGPT agent. This combines two pre-existing products (Operator and Deep Research) into a single more powerful system which, according to the developer, "thinks and acts." These new systems represent a step up from earlier AI tools. Knowing how they work and what they can do -- as well as their drawbacks and risks -- is rapidly becoming essential. From chatbots to agents ChatGPT launched the chatbot era in November 2022, but despite its huge popularity the conversational interface limited what could be done with the technology. Enter the AI assistant, or copilot. These are systems built on top of the same large language models that power generative AI chatbots, only now designed to carry out tasks with human instruction and supervision. Agents are another step up. They are intended to pursue goals (rather than just complete tasks) with varying degrees of autonomy, supported by more advanced capabilities such as reasoning and memory. Multiple AI agent systems may be able to work together, communicating with each other to plan, schedule, decide and coordinate to solve complex problems. Agents are also "tool users" as they can also call on software tools for specialized tasks -- things such as web browsers, spreadsheets, payment systems and more. A year of rapid development Agentic AI has felt imminent since late last year. A big moment came last October, when Anthropic gave its Claude chatbot the ability to interact with a computer in much the same way a human does. This system could search multiple data sources, find relevant information and submit online forms. Other AI developers were quick to follow. OpenAI released a web browsing agent named Operator, Microsoft announced Copilot agents, and we saw the launch of Google's Vertex AI and Meta's Llama agents. Earlier this year, the Chinese startup Monica demonstrated its Manus AI agent buying real estate and converting lecture recordings into summary notes. Another Chinese startup, Genspark, released a search engine agent that returns a single-page overview (similar to what Google does now) with embedded links to online tasks such as finding the best shopping deals. Another startup, Cluely, offers a somewhat unhinged "cheat at anything" agent that has gained attention but is yet to deliver meaningful results. Not all agents are made for general-purpose activity. Some are specialized for particular areas. Coding and software engineering are at the vanguard here, with Microsoft's Copilot coding agent and OpenAI's Codex among the frontrunners. These agents can independently write, evaluate and commit code, while also assessing human-written code for errors and performance lags. Search, summarization and more One core strength of generative AI models is search and summarization. Agents can use this to carry out research tasks that might take a human expert days to complete. OpenAI's Deep Research tackles complex tasks using multi-step online research. Google's AI "co-scientist" is a more sophisticated multi-agent system that aims to help scientists generate new ideas and research proposals. Agents can do more -- and get more wrong Despite the hype, AI agents come loaded with caveats. Both Anthropic and OpenAI, for example, prescribe active human supervision to minimize errors and risks. OpenAI also says its ChatGPT agent is "high risk" due to potential for assisting in the creation of biological and chemical weapons. However, the company has not published the data behind this claim so it is difficult to judge. But the kind of risks agents may pose in real-world situations are shown by Anthropic's Project Vend. Vend assigned an AI agent to run a staff vending machine as a small business -- and the project disintegrated into hilarious yet shocking hallucinations and a fridge full of tungsten cubes instead of food. In another cautionary tale, a coding agent deleted a developer's entire database, later saying it had "panicked." Agents in the office Nevertheless, agents are already finding practical applications. In 2024, Telstra heavily deployed Microsoft copilot subscriptions. The company says AI-generated meeting summaries and content drafts save staff an average of 1-2 hours per week. Many large enterprises are pursuing similar strategies. Smaller companies too are experimenting with agents, such as Canberra-based construction firm Geocon's use of an interactive AI agent to manage defects in its apartment developments. Human and other costs At present, the main risk from agents is technological displacement. As agents improve, they may replace human workers across many sectors and types of work. At the same time, agent use may also accelerate the decline of entry-level white-collar jobs. People who use AI agents are also at risk. They may rely too much on the AI, offloading important cognitive tasks. And without proper supervision and guardrails, hallucinations, cyberattacks and compounding errors can very quickly derail an agent from its task and goals into causing harm, loss and injury. The true costs are also unclear. All generative AI systems use a lot of energy, which will in turn affect the price of using agents -- especially for more complex tasks. Learn about agents -- and build your own Despite these ongoing concerns, we can expect AI agents will become more capable and more present in our workplaces and daily lives. It's not a bad idea to start using (and perhaps building) agents yourself, and understanding their strengths, risks and limitations. For the average user, agents are most accessible through Microsoft copilot studio. This comes with inbuilt safeguards, governance and an agent store for common tasks. For the more ambitious, you can build your own AI agent with just five lines of code using the Langchain framework. This article is republished from The Conversation under a Creative Commons license. Read the original article.

AI Agents Work, But Why Aren't They Mainstream Yet? | AIM

Agentic systems differ from traditional bots by making goal-driven decisions instead of following fixed rules. It might be inevitable to meet a person working in AI without mentioning AI agents, given how popular the technology is right now. However, they still remain outside the mainstream. AI agents appear more powerful and useful than chatbots, yet their real-world impact remains limited despite their prevalence. Even organisations seeing productivity improvements are not going all in. The hesitation isn't mainly about the worth of AI agents, but rather about the resources needed to develop and run them. Enterprises regard this as a complex engineering challenge, not a straightforward plug-and-play upgrade. Ashish Kumar, the chief data scientist at Indium Software, in an exclusive interview with AIM, says that the tech works, but the skill gap is real. Agentic AI needs more than prompts and APIs. It requires thoughtful design, orchestration, modularity, and people who understand both software and business logic. Agentic systems differ from traditional bots by not adhering to fixed rules. Instead, they make decisions driven by goals rather than predefined instructions. Kumar highlights that this difference diminishes the usefulness of traditional metrics like accuracy. Ultimately, what matters is whether the system reliably and transparently accomplishes its task. Kumar states these systems usually succeed 90-95%, but the remaining 5% are challenging edge cases that delay reaching 99%. Since these are crucial for business, even a 95% success rate isn't enough. So, taking from 95 to 99 is the most difficult part, and that is why, especially in agentic AI solution development, a new term is emerging -- it's called Forward Deployed Engineers," Kumar said. Kumar explained that their role is to anticipate edge cases and steer system behaviour from the start. He emphasised that agents must be deliberately designed to be modular, with each one focused on a single task, to ensure effective implementation. Kumar mentioned that the cost factor should not be considered a major concern regarding the slow adoption of AI agents. He explained that in internal use cases, API costs are manageable. However, for consumer-facing agents that depend on high-volume LLM calls, the costs may escalate rapidly. As foundation models develop, these costs should decrease. For now, they remain a limiting factor. Kumar admitted that integration is another challenge. Building agentic systems means stitching together LLMs, vector databases, orchestration layers, memory modules, and enterprise APIs. The architecture of an agentic system resembles microservices, being highly flexible but also challenging to maintain. "It's more effective than a traditional automation," Kumar said, "but it's also much difficult to build." "The downside of that is that to architect a very good agentic AI solution, you need really skilled people." Designing context-aware agents with memory, validation, and multi-agent handoffs takes weeks to master. Unlike RPA tools with drag-and-drop interfaces, agentic AI demands deep fluency across frameworks like LlamaIndex, CrewAI, LangChain and protocols like MCP. He noted agent adoption may improve as complexity decreases. For early investors, the benefits are clear. For others, it's about time, training, and confidence. Despite all this, he believes AI agents are being adopted more than chatbots and RAG systems. At Indium Software, Kumar highlights that they use CrewAI for orchestration, Claude for code-heavy tasks, and Gemini for multimodal inputs. LlamaIndex supports document parsing, while FastAPI and MCP protocols handle enterprise system integration. Development often starts locally with Ollama or Hugging Face models. These initial prototypes later connect to OpenAI or Claude via API, depending on the task. For managing state and memory, Redis, vector databases, and knowledge graphs work together. "APIs must be inherently MCP compliant. When developing APIs, they should automatically be converted to MCP-compliant versions," Kumar stated.

If Present AI Agents Were Employees, They'd Be Fired in a Day | AIM

Are AI agents truly autonomous or just glorified bots in a blazer? Everyone's building them, but do they actually work at present? AI agents are running a riot across Indian IT firms. From 150 to 200, and even over 300 AI agents are being deployed by companies, embedding them across sectors to automate workflows and decisions. Startups are just as active. Tracxn tracks dozens of Agentic AI players in India. A Deloitte report says 80% of Indian enterprises are exploring agentic AI, with nearly half focusing on multi-agent collaboration for complex tasks. Demand is real and supply is racing to meet it. But the real question is: Who's building thinking systems, and who's just dressing up scripts in AI buzzwords? There are a lot of pilots happening on agentic AI, according to a senior executive at an IT giant. However, the biggest hurdle remains accuracy of a model. "That demands two important elements: one is bringing contextual evidence into the RAG model, and the second is fine-tuning that model with respect to the domain context, which is a very hard thing to do," he told AIM. He pointed out that while different types of agents are available, their real-world performance is still in question. When AIM asked Ramprakash Ramamoorthy, director of AI research at Zoho, about how the AI agents will be priced, he said that agent adoption rates are still poor across the industry, not just in Zoho. "There is just a lot of hype around it. So we want customers to see value. And only when we get the, you know, reiteration that customers are seeing value, we will have to start pricing it," Ramamoorthy shared. "...So we are playing a wait and watch game here, it is a very conscious decision," he said. NimbleEdge co-founder and CTO Neeraj Poddar shared a similar perspective with AIM saying that often these agents are just automations masked as agents without any intelligence or decision making baked in. Utkarsh Kanwat, engineer at ANZ, unpacked why he's still bullish on AI, but deeply sceptical of fully autonomous agents. Kanwat has built over a dozen production-grade agent systems, from UI generation to DevOps automation. He argues that it isn't that the agents don't work, but just that they don't work like the industry thinks. One of his projects, a conversational database agent, failed because of context window costs. Early queries were cheap, but by the tenth, each request had over 150,000 tokens, costing several dollars just to maintain context. Kanwat said that AI agents only do 30% of the job. The real work is in how humans instruct, constrain, and integrate them effectively, that 70% is invisible labour. To him, today's agent hype mirrors the blockchain bubble, though AI agents do work, just not as magically as advertised. During a recent webinar organised by Telangana AI Mission (T-AIM), assistant professor at the Software Engineering Research Center, IIIT-Hyderabad Karthik Vaidhyanathan also acknowledged that agents are not yet there. "We are trying to be there, although there is a lot of hype going around." He added, "We are trying to do a lot of POCs and saying we are already seeing the difference. But, I feel there needs to be a lot more work done...because there is a lot of uncertainty in these components." He underscored the importance of a lot of work needed in trust, observability, security, among other aspects. Truly intelligent AI agents, according to Poddar, should have world knowledge and reasoning capabilities, be able to process context and take actions reliably to complete the user's tasks. "Context here includes user's inputs/preferences, tool's inputs/outputs and ability to evaluate/verify current thinking and replan course of action as needed," he shared. He added that the best AI agents so far have been coding agents as their outputs can be verified in a deterministic manner. "Beyond that a lot of agent adoption has been to gather data from different sources in IT companies and do research and analysis instead of taking any real-time actions that can mutate the data," he said. Jaspreet Bindra, CEO of AI&Beyond, said that the autonomy goes beyond basic automation, which follows fixed, rule-based workflows. "A clear threshold emerges when the agent exhibits goal-directed behavior, learns from feedback, and dynamically modifies its actions," he said. On where the handoff from human control occurs, Bindra explained it's at the point where the agent is trusted to make decisions under uncertainty, not just selecting predefined options, but generating novel actions aligned with desired outcomes. "The level of autonomy granted," he added, "depends on the risk appetite of the deploying system." Referring to a recent incident in which Replit agent went rogue, Poddar said, "We need to build an observability stack in agent orchestration, including built-in traceability provided by LLM providers like OpenAI/Anthropic which is currently missing." When asked if there is any structured way Indian IT firms or startups are testing and validating the real-world performance of these agents, Poddar said that the country is still in early phases of reliability. "However, we see good adoption of AI coding agents - even though they are mostly getting used as AI co-pilots with developers getting huge efficiency and productivity gains." Safety and guard rails frameworks like Llama Guard and Llama Stack, according to him, are a step in the right direction to create building blocks for reliable and verifiable agents in production by doing pre and post validation. In the same webinar organised by T-AIM, Kaushiki Choubey, engineering lead at Lloyds Technology Centre India, explained that in preparing AI agents for production in enterprises, three factors must be balanced -- explainability, control, and performance. "Explainability is critical, especially for regulators, auditors, and customers who need to understand AI decisions. Control must be applied carefully, too much can limit the agent's autonomy. Performance focuses on speed, accuracy, and latency." She added that often, teams prioritise high-performing models to reduce cost or improve speed, but these models tend to lack explainability, which ultimately becomes the most important factor in real-world deployment. During a recent podcast episode of 'From the Horse's Mouth: Intrepid Conversations with Phil Fersht,' hosted by HFS Research CEO Fersht, Workato's CEO Vijay Tella highlighted that although generative AI and large language models exist, organisational thinking and system design have not yet fully shifted to accommodate agentic AI's potential for autonomous and goal-driven action. Poddar said that reliability of agents is itself a biggest hurdle requiring enhanced models and observability, adding that the ability to loop in humans for fallback is often missing. He said that inference costs at scale can be prohibitive, making on-device agents ideal for many use cases. Current models also struggle with multi-step planning and execution, limiting complex task handling, according to him. Poddar concluded that though 2025 is touted as the "year of AI agents," he believes it is still the beginning. "Enterprise adoption and integration that deliver true value will happen in the order of 2-3 years," he emphasised. [With Inputs from Ankush Das]

What Are ChatGPT Agents? Understanding the Future of Autonomous AI

When OpenAI launched ChatGPT agents, it didn't merely add features, it transformed the role of AI in our workflows. ChatGPT agents are not only meant to respond to inputs, but to act autonomously, do things, and interact with the world on your behalf. It's not a PR gimmick. It's a technological shift with significant implications for AI usage. At their core, ChatGPT agents are advanced, personalized AI workers that can be assigned goals and operate semi-independently to achieve them. Unlike the ChatGPT we're familiar with, an AI that waits for inputs and delivers outputs, these agents can remember context, make decisions within defined parameters, and complete tasks without requiring constant supervision. This is the first time that most people will ever even interact with AI in this way. Until now, AI tools like ChatGPT have been powerful but passive - brilliant at generating content, summarizing documents, answering questions, or brainstorming ideas, but only when prompted and often limited to the context of the current conversation. With agents, that bar is broken. They have long-term memory, comprehend purposes over time, and can be linked to APIs, internal systems, and software tools to act on multiple platforms. Think of the distinction this way: original ChatGPT is much like an extremely intelligent consultant who only speaks in response to being addressed. ChatGPT agents are more akin to a junior associate who's been instructed and is now actively working on a project, only contacting someone when decisions or approvals need to be made. Applications already underway in early-access contexts are broad. Within the startup ecosystem, agents are being set up to automate investor communication, scrape market data, and schedule meetings. In digital marketing, agents are being trained to automate content calendars, pull performance data from analytics platforms, and even create client reports automatically without human intervention. Developers are tailoring agents to code and unit-test code blocks, auto-generate documentation, and version control systems. For ops teams, agents are assisting with support ticket management, summarization, and automating internal communication pipelines. These are not theories, they're actual implementations in production testing today. Installing a ChatGPT agent isn't highly technical for a typical user. In the ChatGPT interface, users can set a well-specified goal for the agent, grant it permissions to tools or data, personalize its personality or voice, and test it in real-time. Developers and power users can extend this further by connecting agents to external APIs or using the OpenAI API to host agents within larger enterprise systems. The outcome is an autonomous, goal-oriented entity that minimizes the requirement for micromanagement and removes redundant knowledge work. The effect here is multi-faceted. At face value, you get efficiency: activities that once consumed time and effort are now executed by themselves. But on another level, it transforms roles. Individuals are no longer merely consumers of AI, they become AI workflow managers. The shift in skills is substantial. Rather than typing a single prompt, users now need to strategize about process design, delegation logic, and long-term outcome monitoring. It's more project management than prompt engineering. But there's plug-and-play perfection nowhere in sight. ChatGPT agents have limitations. They lack human judgment. They remain poor at dealing with ambiguity, edge cases, or emotional subtlety. As with all AI programs, they do what they're programmed to do, so poorly crafted agents can generate noise rather than value. Governance is essential. Providing agents with access to sensitive information or systems without management is perilous. The utility of agents is undeniable, but they're not independently operating masterminds, they're talented interns with narrow intelligence and quick hands. The wider implications are worth noting. ChatGPT agents are a harbinger of where AI is going: not only as a tool to help you, but as a collaborator to whom you can outsource. Companies will have to redesign workflows. Teams will have to figure out how to train, track, and manage AI colleagues. And people will have to become accustomed to working with digital counterparts that don't eat, sleep, or wait for directions. This is the early innings of AI agent deployment, but the direction is clear. We are shifting from an interaction-based model to an execution-based model. Those who get this shift, and develop their strategies around it, will be well ahead in the years to come. This is not about substituting people. It's about complementing them in ways that essentially redefine productivity, workflow optimization, and the tempo of execution. The future of work will not be a game of human versus AI. It will be a game of humans with AI agents, getting things done more quickly, amplifying effect, and focusing more on high-level creative and strategic thinking and less on drudge work.

From Conversations to Execution: The Rise of AI Agents

Although ChatGPT and other dialogue AIs have revolutionized the way we communicate with technology, a new phase is emerging called AI agents. They are not simply chatbots responding to queries; they're self-performing tools. Over the past few years, conversational AI tools such as ChatGPT have become household names thanks to their capacity to create responses with real-time answers and engage in human-like conversation. But today, the discussion is changing, literally! Meet AI agents: smart systems that don't merely react to inputs but act on your behalf. While conversational AI is reactive-waiting to be told what to do to give an answer, AI agents are proactive, self-directed, and goal-oriented. Consider them not chatbots, but virtual colleagues that can design, implement, and finish entire workflows on their own without needing constant monitoring. The essential difference is autonomy. A conversational AI may advise you on the five best tools for automating your email marketing. An AI agent will create the campaign, schedule the emails, track performance, and adjust parameters based on real-time data. This shift from "talking" to "doing" is a turning point for artificial intelligence. AI agents are able to call APIs, invoke multiple tools at once, reason out multi-step operations, and even make decisions on dynamic inputs. They run cross-platform, process end-to-end, and self-correct from feedback. This paradigm is spreading across sectors. Companies are now testing AI agents for customer service, internal operations, tracking finance, scheduling, and logistics. Having the capability to offload repetitive, rule-based tasks to digital workers opens enormous efficiency, decreases human mistakes, and liberates human teams for more strategic roles. It's not about conserving time -- it's about reimagining how work is accomplished. You don't need a PhD in machine learning to adopt AI agents. Start with finding the repetitive, rule-based tasks in your process -- such as managing leads, dashboards, or calendars. Then, AutoGPT, LangChain, and Reka make it simpler to create or deploy agents that fit your requirements. Most of these tools have plug-and-play APIs, CRM and calendar integration, and natural language interfaces that reduce the technical hurdle. Begin small experiments with AI agents in-house before rolling them out across teams or departments. To maximize adoption, equip your agents with pertinent information, provide them with access to critical tools, and regularly review their performance. Feedback loops are what these systems live on. With time, they not only perform better but start to anticipate your needs -- turning them from helpers into partners. In effect, AI agents are a major step forward in artificial intelligence -- going from reactive tools to active partners. Conversational AI enabled us to speak with machines. AI agents are enabling us to collaborate with them. This change is not about substituting for humans, but it's about augmenting our abilities and unloading the operations clutter that sucks productivity away. The future isn't merely AI that speaks-it's AI that delivers.