WHAT ARE AI 'AGENTS' AND SHOULD YOU CARE?

Introduction

You can't swing a cat without hitting a company claiming they've built "AI agents" these days. Microsoft's got them. Google's got them. Every startup that was selling chatbots last month is now flogging "autonomous AI agents."

The marketing suggests we're on the verge of AI employees that'll handle your emails, book your meetings, and probably make you a cup of tea. The reality's considerably less exciting, though there are some genuinely interesting developments buried under the hype.

What they actually mean

An "AI agent" is supposed to be AI that can perform multi-step tasks without you holding its hand through every click. Unlike a chatbot that just answers questions, an agent supposedly takes action and makes decisions about what to do next.

The classic example is scheduling. You tell it to arrange a meeting, and it handles the back-and-forth emails, checks everyone's calendars, adjusts when there are conflicts, and sends the invite. The system decides what steps to take based on the responses it gets.

Here's what most companies aren't telling you clearly. What they're calling "AI agents" spans a huge range of capability, from barely more than automation with a chat interface, to systems that genuinely can plan and adapt their approach.

Agent versus agentic AI

There's a distinction worth understanding, even though companies use these terms interchangeably.

An "AI agent" in marketing speak often just means any AI system that does things rather than just answering questions. Your bank's system that can check your balance? They'll call that an agent.

"Agentic AI" is supposed to mean something more specific. Systems that can break down goals into steps, choose which tools to use, monitor whether things worked, and revise their approach when they don't. Think of it as the difference between following a recipe exactly versus actually cooking, where you taste as you go and adjust.

Most products sold as "AI agents" don't actually have agentic capabilities in any meaningful sense. They're following scripts, just more flexible scripts than old automation.

What's genuinely new here

There is something architecturally different about proper agentic systems, even if it's not the revolution the marketing claims.

Traditional automation required you to map out every step in advance. If this happens, do that. If that fails, do this other thing. Every possibility had to be coded explicitly.

Modern agentic systems can figure out steps as they go. You give them a goal, and they decompose it into tasks, decide which tools to use, check if each step worked, and adjust their plan based on what happened. That's genuinely different from automation, even if the end result sometimes looks similar.

They also handle tool selection differently. Old automation was hard coded. Step one, query this database. Step two, send this email. Agentic systems decide at runtime which tool to use based on the situation.

The third difference is stateful iteration. These systems can observe outcomes, realise something didn't work, try a different approach, or escalate to a human. That's closer to how control systems work than how traditional scripts operate.

What it looks like in practice

Some genuine examples exist where these capabilities matter.

Scheduling assistants like x.ai can handle the coordination because they can adapt their approach based on responses. If someone says they're only free Thursdays, the system adjusts its strategy without being explicitly programmed for that scenario.

Customer service bots with proper agentic capabilities can navigate unexpected situations better than scripted systems. If a refund request doesn't match standard patterns, a decent agentic system might check order history, review return policies, and determine an appropriate response, rather than just giving up and transferring to a human.

The key is whether the system can genuinely plan, monitor, and adapt, or whether it's just following more flexible scripts.

The marketing problem

Most of what's being sold as "AI agents" is just normal automation with a chatbot stuck on top, without any of the agentic capabilities described above.

Your bank's "AI agent" that can check your balance? That's a chatbot querying a database. There's no planning happening. No tool selection under uncertainty. No adaptation based on outcomes. It's perfectly good technology, but it's not what the term implies.

The genuinely new bit in these systems is the natural language understanding. Being able to ask questions normally instead of navigating menus is a real improvement. But companies are acting like they've invented something world changing when they've just made their existing systems easier to talk to.

Where proper agentic systems work

When you actually have systems with planning, tool selection, and adaptation capabilities, they handle certain categories of tasks reasonably well.

Narrow, bounded domains with clear success conditions. Scheduling works because the goal is well defined and there are limited variables. The system can try different approaches and know whether it succeeded.

Tasks with structured information and clear rules. Processing invoices and updating records can work because the system can check its work and retry if something's wrong.

Situations where the system can fail safely and learn from feedback. If a first attempt doesn't work, try something else. This only works when failure isn't catastrophic.

Where agentic systems fail

Anything requiring judgment beyond the immediate task.

Your scheduling agent might find a time everyone's free, but it doesn't know the context. Like "Sarah hates early meetings after long flights". And you definitely don't want anything scheduled Friday afternoon before a bank holiday! These require human judgment that goes beyond optimising calendar availability.

Novel situations remain the killer. Agentic systems work by decomposing goals using patterns they've learned. Throw them something genuinely unprecedented and they'll either refuse to help or confidently do the wrong thing.

High stakes decisions where errors matter. Current agentic systems are fragile. They make mistakes. They misunderstand context. You don't want them handling anything where getting it wrong has serious consequences.

The architecture they're not explaining

Proper agentic systems are built around a set of capabilities that work together to enable genuine autonomy, even if limited.

Goal representation is about understanding what you actually want to achieve, not just processing the literal words you said. When you tell a system "sort out my travel," a proper agentic system needs to understand that means booking flights, arranging accommodation, checking visa requirements, and coordinating timing, not just searching for flight prices.

Planning and task decomposition is how the system breaks that goal down into concrete steps it can actually execute. It needs to figure out the sequence, dependencies, and what to do if any step fails.

Memory is crucial and expensive. Short term memory means the system remembers context from earlier in the conversation. Long term memory means it can recall your preferences and patterns from previous interactions. This is extremely difficult to achieve well and costs significant computing resources, which is why most commercial products don't bother.

Feedback loops allow the system to monitor whether each step actually worked, rather than blindly executing a sequence and hoping for the best. The system needs to check results and understand when something's gone wrong.

Error recovery means when something fails, the system can try different approaches instead of just giving up. If the first hotel is fully booked, try alternatives. If the cheapest flight doesn't work with your schedule, recalculate.

Escalation logic helps the system recognise when it's out of its depth and needs to hand over to a human rather than continuing to make things worse. An AI system knowing when it doesn't know something (and actually recognising that fact!) is the holy grail of AI in general.

When all these components work together, you get something that can genuinely pursue goals with limited human oversight. When any are missing, you get systems that look impressive in demos but fail unpredictably in practice.

The privacy trade-off

For an agentic system to actually work, it needs deep access to your information and your systems, far beyond what a simple chatbot requires.

A scheduling agent needs to read your emails, access your calendar, see who you communicate with regularly, and understand your preferences and patterns. The more sophisticated the agentic capabilities, the more invasive the access becomes. Systems with memory need to store information about your behaviour over time. Systems that can act across multiple tools need authentication to all those services, which means they're essentially operating with your identity across your digital life.

Think about what that actually means. You're giving software permission to read your private communications, understand your relationships and habits, make decisions on your behalf, and potentially remember all of this indefinitely. That's not like letting a chatbot answer questions from a knowledge base. That's handing over a significant portion of your digital autonomy to a system that will inevitably make mistakes, might be compromised, and is certainly collecting data about you.

This isn't automatically bad, but pretending it's not a significant privacy concern is dishonest. A proper agentic system isn't just software you talk to. It's software acting as you, with access to information and systems that define your private life, and the more capable it becomes, the more access it demands.

The reality check

The honest assessment is this: there's a meaningful architectural shift happening, and agentic AI systems that can plan, adapt, and iterate are genuinely different from traditional automation in ways that enable some new capabilities.

But these systems are fragile, expensive to operate, limited to narrow domains, and years away from the autonomous AI employees that the marketing promises, and most products labelled "AI agents" don't actually have these capabilities anyway. What we've got right now is mostly tool-using chatbots with varying degrees of sophistication, where some can do basic planning and adaptation, most can't, and almost none have the full architecture that would justify calling them properly agentic.

The technology is real and the direction is genuine, but the hype is still wildly premature.

Should you care

Ignore the hype about autonomous AI transforming everything. Do pay attention to whether specific applications actually solve problems you have, and whether they demonstrate genuine agentic capabilities or just flexible automation.

If you spend hours coordinating meeting times, a decent scheduling agent with proper planning and adaptation might genuinely save you hassle. If you run a shop and most customer queries follow patterns, a system that can handle variations and adapt its approach could free up time for complicated cases.

But approach any claims about AI agents with serious scepticism. Ask what the system actually does. Can it plan and adapt its approach, or is it following scripts? What access does it need? What happens when it gets things wrong? How does it handle situations it hasn't seen before? Also, ask what they were calling this technology last year, because there's a good chance they've just renamed their existing automation and added "agentic" to the marketing materials.

The useful applications of agentic AI exist and the technology is developing in interesting directions. But we're nowhere near the autonomous AI revolution the companies selling these products would like you to believe. The gap between current reality and marketing promises is enormous, and will likely remain so for years.

Browse all topics → Index