Perspectives

This new capability emerges from the convergence of several technical advances. The underlying language models have become more capable at planning and reasoning. Frameworks for tool use and function calling have matured, allowing AI systems to reliably interact with APIs, databases, and software interfaces. Persistent context management systems enable these agents to maintain context across long interactions and even between sessions. It is also important that techniques for constraining and validating agent behaviour have improved enough that companies are willing to deploy these systems in production environments where errors have real consequences. 

We are entering the era of agentic AI, and it represents a very real change in how AI systems interact with the world. If ChatGPT in November 2022 represented an inflection point, the conversation around artificial intelligence shifted in again 2025 with the focus turning to agentic AI. Where large language models dominated headlines with their ability to generate text and answer questions, agentic AI is something more consequential: AI systems that don’t just respond but which can act.

The progression from traditional AI to agentic systems follows a logical evolution. Early AI systems were pattern driven – they could classify images, transcribe speech, or predict outcomes based on historical data. Large language models took this further by demonstrating reasoning capabilities and the ability to engage in seemingly human dialogue. But these systems remained fundamentally reactive. They waited for human input, generated a response, and then waited again.

Agentic AI breaks this pattern. Agentic systems can pursue goals over extended timeframes, break down complex objectives into subtasks, use tools to interact with their environment, and adapt their strategies based on feedback. An agentic system doesn’t just tell you how to book a flight – it can search flight options, compare prices, make a reservation, and send you a confirmation, all while handling unexpected complications like sold-out flights or schedule changes. 

The economic drivers are equally compelling. Hiring knowledge workers who can perform complex, multi-step tasks is difficult and agentic AI offers a possible path to overcome this limitation. A customer service agent that can resolve issues end-to-end, a research assistant that can gather and synthesize information from dozens of sources, or a software development agent that can implement features from specification to deployment – these capabilities promise productivity gains that go beyond incremental efficiency improvements.

The technical obstacles to reliable agentic AI are formidable and they are more difficult than they initially appear. The first and perhaps most obvious challenge is what researchers call the “action space problem.” While language models can generate text reliably, taking actions in the world requires interacting with systems that weren’t designed for AI agents. APIs change without notice. Websites update their structure. Software interfaces vary across platforms. An agent that worked perfectly yesterday may fail today because a service provider modified their authentication flow.

Reliability presents an even deeper challenge. When a language model generates an incorrect response to a question, the consequences are typically limited. When an agentic system takes an incorrect action—sending an email to the wrong person, deleting important files, or making an unauthorised purchase—the consequences can be far-reaching. The probabilistic nature of language models means they will occasionally make mistakes, but the systems we integrate them with often assume deterministic, correct behaviour. This mismatch creates fundamental tension.

Another difficult challenge is known as the planning problem. Simple tasks with clear success criteria can be automated effectively, but many real-world objectives are ambiguous and require the navigation of trade-offs. When multiple approaches might work, how should the agent choose? Human judgment often involves intuition and contextual understanding that are hard to codify. Training agents when to be creative and when to follow established procedures is an unsolved problem in the field of agentic AI. 

When humans encounter unexpected situations, we naturally adapt our approach, seek clarification, or escalate to someone with more authority or expertise. Agents tend to either plow ahead with flawed assumptions or give up entirely. Agentic systems need to recognise when they are out of their depth and gracefully hand off to human oversight. This requires AI agents to communicate uncertainty effectively which in turn requires advances in both the models themselves and the frameworks surrounding them.

The security implications deserve particular attention. Agentic AI systems represent a significant expansion of the attack surface. Traditional software has well-defined inputs and outputs. Agents can potentially interact with any system they have access to, interpret natural language instructions, and make autonomous decisions. This creates opportunities for prompt injection attacks, social engineering, credential theft, and unauthorised access that go beyond traditional cybersecurity concerns. An attacker who can influence an agent’s goals or its perception of the environment can cause it to take arbitrary or unexpected actions.

Then there’s the question of verification and testing. How do you validate that an agent will behave correctly across the virtually infinite space of possible scenarios it might encounter? Traditional software testing relies on deterministic behaviour and comprehensive test suites, but agents operating in open-ended environments resist this approach. Formal verification methods that work for traditional software don’t readily extend to systems whose behaviour emerges from probability distributions over billions of parameters.

The technical challenges are substantial but they may prove easier to solve than the societal ones. Agentic AI requires us to confront questions about accountability, autonomy, and the nature of work that we’ve largely left unanswered in the era of AI assistants.

Accountability is a thorny issue. When an agent makes a consequential mistake, who bears responsibility? The company that deployed it? The developers who built it? The AI provider whose model powers it? The user who gave it instructions? The traditional legal frameworks we use to assign liability assume human decision-makers whose actions can be evaluated against standards of competence and reasonableness. Agents occupy an awkward middle ground – capable enough to act autonomously, but not conscious beings who can be held morally accountable for their choices.

This problem intensifies as agents gain more autonomy. A system that simply executes clear human instructions represents a straightforward case of a tool. But agents that interpret vague objectives, make judgment calls about how to achieve them, and potentially operate for extended periods without human oversight begin to resemble something closer to employees or contractors. Our legal and organisational frameworks don’t have clear categories for entities that occupy this space.

The societal implications are far-reaching and employment market impacts are never far away when considering the impact of AI. Some job roles will be transformed or eliminated – junior analysts conducting research, customer service representatives handling routine inquiries, administrative assistants scheduling meetings and managing logistics. But the second-order effects may prove more significant. If agents can handle routine work, what happens to career progression paths that rely on junior employees learning through experience? How do organisations maintain institutional knowledge if experienced workers aren’t training successors? What becomes of the social fabric in workplaces when interactions become human-to-agent rather than human-to-human?

Building successful agentic AI systems requires enormous resources—computing infrastructure, training data, specialised talent, and access to the tools and APIs that agents interact with. This creates natural advantages for large technology companies and raises concerns about whether we’re building toward a future where a handful of corporations control the infrastructure through which work, commerce and communication flows. The dynamics differ from previous waves of automation because agents can operate across such a broad range of domains.

Privacy takes on new dimensions when agents act on our behalf. An AI assistant that can read our email, access our calendar, make purchases, and communicate with others has unprecedented insight into our lives. Even if the agent itself doesn’t retain this information permanently, the systems it interacts with create audit trails that could be subpoenaed, hacked, or used in ways we didn’t anticipate. The aggregation of information across multiple agents and services creates opportunities for inference and surveillance that go beyond what each individual interaction reveals.

There’s also the question of what is known as cognitive autonomy. As agents become more capable at planning, decision-making, and execution, there’s a risk that humans increasingly defer to their recommendations without engaging critical thinking that these decisions warrant. This isn’t a new phenomenon—we’ve seen similar changes with GPS navigation, search engines, and recommendation algorithms. But agents that can act across multiple domains and pursue complex goals could accelerate this trend.

Perhaps the most contentious debate surrounding agentic AI concerns the necessity of human oversight. Should there always be a human in the loop, reviewing and approving agent actions? Or should we allow agents to operate autonomously, potentially even interacting with other agents without human involvement?

The case for mandatory human oversight rests on several foundations. First, there’s the fundamental unreliability of current systems. Language models hallucinate, misunderstand context, and make errors that human can identify and address. Requiring human review provides a crucial safety valve. Second, many decisions involve value judgments, trade-offs, and contextual considerations that resist algorithmic treatment. A human can recognise when a situation calls for flexibility, creativity, or empathy in ways that AI systems cannot. Third, human oversight preserves accountability. When humans review and approve agent actions, the chain of responsibility remains clear.

The practical challenges with universal human-in-the-loop approaches become apparent quickly. If every action requires human approval, we’ve essentially built an expensive task recommendation system rather than an agent. The promise of agentic AI – the ability to delegate complex, multi-step workflows – depends on agents being able to act without constant supervision. This suggests a need for graduated autonomy, where agents can handle routine, low-risk actions independently but must seek approval for more important consequential decisions.

There’s an argument that human-in-the-loop requirements might themselves introduce risks. Humans reviewing agent recommendations are subject to automation bias – the tendency to trust computer-generated suggestions even when they’re flawed. If we create systems where humans nominally approve actions but rarely override agent recommendations, we may get the worst of both worlds: the illusion of human oversight without its benefits, combined with slowed execution and increased costs. 

The financial markets provide both a cautionary tale and a potential model. High-frequency trading algorithms interact with each other millions of times per second, and these interactions sometimes produce unexpected and potentially harmful outcomes—flash crashes, liquidity crises, and market dynamics that don’t reflect fundamental value. But the markets have also developed mechanisms for controlling these systems: circuit breakers, kill switches, monitoring for abnormal patterns, and strict liability for losses. Similar frameworks might apply to agentic AI interactions.

The agent-to-agent interaction scenario raises the stakes considerably. Imagine two companies, each using agentic AI systems to handle procurement and sales respectively. These agents could potentially negotiate contracts, agree on terms, exchange information, and finalise transactions without any human involvement. The efficiency gains would be enormous, but so would the risks. How do we ensure these agents are actually negotiating in their principals’ interests? Could agents, through repeated interactions, effectively collude in ways that disadvantage the humans they ostensibly serve?

The problem is that determining what counts as low-risk or routine is surprisingly difficult. An action that seems innocuous in isolation might have unforeseen consequences when combined with other actions or external circumstances. Email communication provides a useful example. Sending a single, well-crafted email might seem like a low-risk action an agent could handle autonomously. But what if that email creates legal obligations? What if it reaches unintended recipients? What if the tone, while grammatically correct, is inappropriate for the relationship or cultural context?

The autonomy question may ultimately require different answers for different domains and risk levels. Financial transactions above certain thresholds might require human approval. Communication with external parties might need review. But routine internal tasks – scheduling, data organisation, research – might appropriately proceed with minimal oversight. The challenge lies in creating frameworks that can adapt to context while remaining understandable and auditable.

We need better tools for constraining and monitoring agent behaviour. This means advances in formal verification, better frameworks for specifying objectives and constraints, and systems that can explain their reasoning and flag uncertainties. We need clearer legal frameworks that establish liability without stifling innovation and organisational policies that specify when agents can act autonomously and when human oversight is required.

The rise of agentic AI is not a development we can simply opt out of. The competitive pressures are too strong, the potential benefits too significant. Organisations that successfully deploy these systems will gain substantial advantages over those that don’t, but proceeding without careful consideration of the technical limitations and societal implications would be reckless.

Most importantly, we need to resist the temptation toward technological determinism – the assumption that because we can build fully autonomous agents, we should. The question isn’t whether we can eliminate humans from decision-making processes, but rather how we can use these technologies to augment human judgment, preserve meaningful human agency, and create systems that remain accountable to the people they affect.

The next wave is here. How we navigate it will shape not just the future of technology, but the nature of work, the distribution of power, and the preservation of human autonomy in an increasingly automated world. These aren’t questions with easy answers, but they’re questions we can not afford to defer.