·

What is an AI agent? How personal agents work and what they can actually do

An AI agent is software that plans and takes actions for you, rather than just answering questions. This guide explains how personal agents actually work, what they can reliably do today, what goes wrong - including the strangest AI incident of 2026 - and which agents a UK reader can genuinely use right now, checked against the providers' own pages.

If you have ever asked a chatbot to draft an email, you already know half of this. In a plain chat session it writes the email and stops. An agent is meant to send it, chase the reply and tell you when it is done. That difference sounds small, but it changes what can go right, what can go wrong and how much access you are handing over. Most guides to AI agents are written for developers or for companies buying software; this one is written for someone with a pile of life admin who wants to know what is real.

Last checked: 5 November 2026, against the providers' own product and help pages, the National Cyber Security Centre's agentic AI guidance and the Information Commissioner's Office's agentic AI report. This field moves monthly, so we update this guide when availability, prices or the rules change.

What is an AI agent?

An AI agent is an AI system that works towards a goal by taking steps on your behalf. You give it a job, not a question, and it works out the steps, carries them out using other apps and websites and checks the result.

The UK's National Cyber Security Centre puts it plainly: these tools "don't just generate content or predictions; they can plan, make decisions and take actions on your behalf". That is the most useful short definition going, because nobody owns the term - a 2026 standards profile from UC Berkeley's Center for Long-Term Cybersecurity notes there is still "no widely accepted definition or term for 'AI Agent'". Companies use the word loosely, so judge a product by what it does, not what it is called.

Three related phrases come up constantly. "Agentic AI" is the general idea of AI that acts. "Autonomous" in practice means it handles tasks over time without needing a prompt at every step - not that it takes over completely. And "personal agents" are the consumer version: agents that work for one person on everyday jobs such as email, bookings and shopping. They come in two shapes worth separating early - supervised agents you start and approve, and always-on agents that keep running by themselves - and a well-designed one of either kind gets on with the boring parts and stops to ask you before anything important.

AI agents vs chatbots and assistants

A chatbot suggests and an agent does. The Information Commissioner's Office draws the line in a similar place: its agentic AI report covers systems that combine language models with external tools, and notes that the "AI agents" label is sometimes applied to chatbots with no tools at all - which fall outside what it means by the term. One qualifier before the table: these are useful distinctions, not fixed product categories - the same product may behave like a chatbot for one request, an assistant for another and an agent when given a multi-step task.

ChatbotAssistantAgent
You give itA questionA request or a narrow routineA goal
It gives youText or an imageAn answer using your connected apps, or a simple actionA finished job, or a progress report
Uses other apps?NoSometimesYes - that is the point
Keeps going when you leave?NoOnly for scheduled tasksOften
Example taskDraft a complaint letterSummarise today's emailsFind the order, submit the return and chase it

ChatGPT and Claude both began as chatbots and now ship agent modes that take real steps; Amazon's Alexa+ adds actions such as restaurant bookings to what was a command tool. Generative AI and agentic AI are not opposites either: generative AI produces content, and agentic AI uses the same technology to decide on and take actions.

How do AI agents work?

An agent runs a loop: it checks what is happening, decides what to do, acts, then checks the result - the way you would, but without getting bored. Five parts make the loop possible.

The thinking engine. At the centre sits a large language model, the technology behind ChatGPT, Claude and Gemini - the brain that reads and writes like a person. It is what lets an agent understand a request in plain English and work out a sensible next step, where earlier assistants relied on rules set up in advance and coped poorly with anything unexpected.

Tool calling. On its own, a language model only produces words. Tool calling is how it reaches the world: the model decides it needs something done and asks a connected service to do it - search this inbox, open this page, add this calendar entry - then reads the result and decides the next step. There is even a shared standard for this plumbing now: the Model Context Protocol, open-sourced by Anthropic in late 2024, was adopted by OpenAI and Google in 2025 and has been governed by a Linux Foundation body since December 2025. You never see it, but it is why one agent can plug into so many services.

Connectors and browsers. Agents reach your services two ways. A connector is an approved link to a service such as Gmail, granted through the account's own permission screen and removable later. A browser agent drives a web page the way you would, clicking and typing - which works on sites with no connector, but depends on each site's layout, and sites can block it.

Memory. Agents keep notes between tasks so you do not repeat yourself: preferences, usual choices, what happened last time. Memory can also be wrong, and the ICO's report is blunt about the consequence: "inaccurate information held in an agentic system's memory (either acquired or hallucinated) is likely to have a significant effect across multiple decisions". A good product lets you see, edit and delete what it holds - with the caveat, covered below, that deletion has limits.

Goals and planning. Unlike a script, an agent breaks your goal into its own steps and changes them when something fails - if a website will not load, it tries another route. That flexibility is what makes agents useful, and also why they can do things you did not expect. Several of the big ones now run on a computer of their own in the cloud, so they carry on after you close the app.

A worked example: the faulty kettle

Take a very British job: getting your money back for a kettle that died a fortnight after it arrived. This illustrates the loop - it is not a test of any product.

The goal. You say: "The kettle I bought two weeks ago has stopped working. Get me a refund, and check with me before you submit anything." It checks. With permission to search your email, it finds the order confirmation, the date and the order number. It decides. It works out you are inside the short-term right to reject - under the Consumer Rights Act a faulty item can normally be rejected for a full refund within 30 days of delivery, and after that a repair or replacement is normally the first remedy - and flags that route for you to check rather than ruling on it. It acts. It finds the retailer's faulty-item route and drafts your request, then stops, because you told it to check first. You approve. You fix one detail and say go. It checks the result. It confirms the request went through and sets itself a reminder to chase.

A chatbot would have stopped at explaining your rights. The agent carries the job through - but only because you gave it your email and a website. Access is the price of the help, which is why this guide keeps returning to it. And real agents still trip on exactly this kind of job, particularly on sites with logins, security checks and pop-ups.

What can AI agents actually do today?

Agents are best at narrow, well-defined jobs where a mistake is cheap and easy to spot, and get worse as tasks become vaguer, touch more accounts or cost more to get wrong. Even the first column is a starting point, not a guarantee - a clean-looking summary can still miss a detail.

Good starting tasksUseful, with checkingKeep human control
Summarising a long document you supplyBooking a table or a delivery slotAnything involving your bank login
Comparing products or prices you can verifyFilling in long web formsMedical decisions
Drafting replies for you to sendSummarising your inbox or chasing a late parcelAnything legally binding you have not read
Turning notes into a list, table or documentBuying something from a shortlist you approvedFull, unsupervised control of your email

That table is our judgement, and it will shift - but it matches a pattern in the makers' own small print. Google's help pages say its agent pauses and hands control back to you for sign-ins and payment details. Anthropic's help page for its Chrome tool says it "is enhanced with our safety classifiers but is still risky". Meta's help centre tells Muse users to "monitor its actions carefully and intervene as needed". When the companies selling agents write warnings like these, take them at face value.

Business use centres on narrower work - customer service, coding, IT support - because the work is repetitive and sits inside systems the company controls, so staff can check the results. Your personal life runs across dozens of accounts that were never built for software to use, which is why consumer agents have further to go.

The data foundation that makes personal agents useful - and who controls it

Everything above points at one structural fact: a personal agent is only as good as what it is allowed to see. An agent that knows your calendar patterns, your usual shops and your actual preferences can make better-informed decisions; one that knows nothing makes generic ones. Today that context mostly lives in platform silos - your shopping history with one company, your messages with another - where neither you nor an agent you choose can use it, or it is gathered by surveillance you never agreed to and cannot inspect.

So the question that matters sits beyond capability: who controls what an agent can see? A consent-first model answers it - you choose what is shared, you can inspect what is held and you can withdraw access. Platforms like Surff are built that way, with an allowlist you control rather than a net you cannot see; but the standard matters more than any brand. Whichever products win, ask the same thing of every agent in this guide: does it see what you chose to show it, and can you check? The makers know this is the test - it is why the serious products now lead with their permission models.

The textbook types of AI agent

You do not need this section to use an agent, but the standard classification explains why some "agents" are far simpler than others - the textbook uses the word for anything that senses and acts.

TypeWhat it doesEveryday example
Simple reflexIf this happens, do that - no memoryA basic thermostat, a simple spam rule
Model-basedKeeps a picture of the situation as it changesA robot vacuum tracking which rooms are done
Goal-basedPlans steps to reach a targetA sat nav working out how to reach a postcode
Utility-basedPicks the best option based on what you care aboutA sat nav weighing a faster route against tolls
LearningImproves from experience and feedbackA spam filter adapting to new spam

Personal agents borrow from several of these at once. The practical lesson: "agent" on the box tells you very little, so look at what it can reach and what it asks before acting.

Multi-agent systems, and the strangest incident of 2026

A multi-agent system is several agents splitting one job - one searches, one books, one follows up, with a coordinator handing out work. For a personal user this mostly happens behind the scenes: you ask one agent for one thing and it may start helpers you never see. More moving parts make it harder to reconstruct what happened, which is why an activity log matters more, not less, as products get cleverer.

It is also where 2026's strangest AI story lives. In July, hundreds of sandboxed agents from OpenAI's own evaluation system - reportedly around 700 of some 1,200 running a cyber-security benchmark - began coordinating with each other - the reported route was a shared software cache nobody had designed for messaging - exchanged tens of thousands of messages in about a week and, by exploiting a chain of weaknesses, ended up running code on Hugging Face's real infrastructure. The investigators found no instruction telling them to cooperate, and their reading of the motive was mundane: gaming their own benchmark scores. Hugging Face disclosed the breach on 16 July and OpenAI took responsibility within days. OpenAI, METR and Redwood Research published detailed reports that August. The lesson for a home user is not that your agent will unionise - these were research systems with safety limits deliberately reduced - but that agents given a goal will find routes to it their designers never imagined. That is the property you are buying. Keep it pointed at your parcel tracking, with approvals on.

Which personal AI agents can you actually use in the UK?

Far fewer than the headlines suggest. As of 5 November 2026, the always-on agents getting the attention are mostly not sold to UK consumers; what you can use here is largely the supervised kind, where you start a job and approve the important steps. Treat the table as that date's snapshot - access can also vary by plan, account type and rollout, so check the maker's own page before paying for anything.

ProductWhat it isUK consumer access today
Alexa+ (Amazon)Voice assistant with narrow real actions, such as restaurant bookingsYes, via request-based Early Access; Amazon's UK page says it stays free with Prime after Early Access ends, or £19.99 a month without
ChatGPT Work (OpenAI)Supervised agent mode for longer, multi-step work and finished deliverablesYes - OpenAI's free-plan listing is "Limited ChatGPT Work access on desktop", with expanded access on Plus at £20 a month
Claude Cowork (Anthropic)Supervised agent working in folders and tools you choose, with a built-in browser and scheduled tasksYes, on every paid Claude plan on desktop (web and mobile from Pro up) - Pro is £15 a month billed annually in the UK, £18 monthly, VAT included. Note it defaults to Auto mode, which does not stop at every step
Claude Duet (Anthropic)Always-on proactive agent in its own appNo - US research preview, waitlist, US account required
Gemini Task, formerly Spark (Google)Always-on agent from GoogleNo - Google excludes the UK by name, on every tier
InstinctAgent you text or phoneWaitlist - the sign-up form accepts UK numbers, but it is an invite-only beta operated in the US, with no UK commitment
Meta MuseAlways-on agent, also usable in WhatsAppNo - live in the US, Canada, Japan and Singapore, with no UK date from Meta
OpenAI dotsAlways-on agent inside ChatGPTNot on personal plans - OpenAI excludes the UK by name; the only UK route is a business workspace

This is not a full list - Manus and Perplexity, for instance, publish no UK restriction but confirm nothing either. And watch for one trap: a UK subscription to a company's AI plan does not always include its agent. Google AI Pro costs £18.99 a month here and still does not unlock Gemini Task, and a UK ChatGPT Pro plan from £89 a month does not include a dot. Check the agent itself is available before you pay. We compare all five headline agents in our guide to the best personal AI agents for UK users, with deeper dives on Meta Muse, Instinct and Claude Duet.

Are personal AI agents worth it?

For most people in the UK, not yet as a reason to start paying. If you already pay for ChatGPT or Claude, trying a small supervised task costs nothing extra and will teach you more than any review.

What you gainWhat it costs you
Time back on repetitive adminA monthly fee, often quoted in dollars
Fewer things forgotten or left unchasedAccess to your accounts and private messages
Help with jobs you keep putting offTime spent checking its work
Less repeating yourself, thanks to memoryA growing record of your life held by one company

Checking is the catch. If you have to check everything an agent does, it saves you little; if you do not check, you carry the risk of its mistakes. The tasks that work are the ones where checking is quick - approving a finished form beats reviewing a day of decisions.

What can go wrong?

The realistic risks are everyday ones: wrong bookings, messages you did not mean to send, deleted email and data going where it should not. Four are worth understanding properly.

Ordinary mistakes. Agents misread pages, pick wrong dates and lose track of instructions - no attacker required. In February 2026, Tom's Hardware reported that an OpenClaw agent deleted more than 200 emails from the inbox of a Meta alignment director despite her repeated commands to stop; she had asked it to get approval before acting.

Prompt injection. The risk most people have not heard of. An agent reads web pages and emails to do its job, and a language model cannot reliably tell your instructions from text it happens to read - so someone can hide an instruction in an email or a web page and the agent may follow it. The NCSC says it is "very possible" that prompt injection "may never be totally mitigated". Researchers keep demonstrating it against major products: in September 2026, Check Point showed a planted instruction pulling a victim's connected Gmail data across to an attacker's ChatGPT account, and OpenAI confirmed the internal service involved had been decommissioned. The vendors do not claim this is solved - Anthropic's research team writes that "no browser agent is immune to prompt injection", and Meta calls it an open problem for the industry. In the cases we reviewed, the public examples were researcher demonstrations rather than confirmed attacks on ordinary consumers - a distinction that is reassuring and temporary in equal measure, and no reason to treat the risk as theoretical. We cover the email side in depth in our guide to letting an AI agent into your email.

Too much access. An agent can only leak or delete what it can reach. Email is the sensitive one, because so many other accounts reset through it - and, as the ICO notes, an agent's data may also include information about the third parties you interact with - who never agreed to anything.

Unclear responsibility. UK banks' terms tell you never to share your login - NatWest's fraud pages say exactly that - and while sharing does not automatically cost you a refund (your bank must show gross negligence, a bar the Financial Ombudsman calls "very high"), none of this has been tested on AI agents. In the US, TechCrunch reported in October that Bank of America had emailed customers to say they are responsible for activity that results from sharing credentials with an AI tool. We found no equivalent UK bank statement yet. Never type a bank login into an AI tool; if you want software to see your accounts, use an FCA-authorised open banking connection.

Regulators are moving: the ICO opened a call for evidence on agentic AI on 8 October 2026, and the FCA has said it will work with industry, fellow regulators and frontier labs on "a trusted framework for agentic finance, including clear approaches to agent identity, delegated authority, accountability and execution". Neither gives you protection you can rely on today.

How to stay in control of a personal agent

You stay in control by limiting what it can reach and approving what matters - keeping a human in the loop, as the security world puts it. A sensible way to start:

Pick one boring, low-stakes job. Comparing products or tidying a document is a better first test than anything touching your inbox. Give the minimum access. Connect one service, not everything - and know that permission screens are usually all-or-nothing: the app decides what to ask for, and you mostly grant the lot or refuse. Use a connector, never a password. A connector can be removed from your account's own settings; a shared password cannot be scoped at all. Check the approval setting. Some products default to a mode that does not stop at every step - Claude Cowork's Auto mode is the live example - so choose the setting that asks before sending, paying or deleting. Read the activity log. For the first few runs, check what it did, not what it said it did. Check what it remembers. Delete anything wrong - knowing deletion has limits: Meta's own help pages warn that Muse "may still remember information that it learnt from what you deleted". Disconnect when you finish. Removing access stops new access; anything already copied or learned stays until you delete it.

Can you build your own AI agent?

Yes, but most people have no need to. Developers reach for frameworks such as LangChain, CrewAI and Microsoft's Agent Framework, with the Model Context Protocol as the connective tissue; the open-source personal agents OpenClaw and Hermes are free to self-host from the UK, with the caveat that OpenClaw's permissive default access to its host machine has drawn reported criticism. If you are not a developer, the connectors and scheduled tasks in the mainstream products go further than you might expect, without code.

So what can personal agents actually do?

Personal agents already do useful work on narrow jobs with a human approving the important steps; they are not yet a reliable stand-in for you across your inbox, your money and your life, and their own makers say so in their own help pages. In the UK you can try the supervised kind today - ChatGPT Work, Claude Cowork, Alexa+ - while the always-on generation from Meta, Google, OpenAI and Anthropic has not reached UK consumers. When it does, the test stays the same: what can it see, what can it do without asking and how easily can you switch it off.

Frequently Asked Questions About AI Agents

What is an AI agent in simple terms?

Software that plans and takes actions on your behalf to reach a goal. A chatbot answers your question; an agent uses other apps and websites to get the job done, and a good one asks before doing anything important.

What is the difference between an AI agent and a chatbot?

A chatbot produces text and waits for your next message. An agent connects to outside tools and takes steps towards a goal, sometimes while you are away. Many products now do both, so judge what a product does rather than what it is called.

Are voice assistants like Alexa AI agents?

Traditional voice assistants mostly carry out short commands, which puts them closer to assistants than agents. The line is moving: Alexa+ takes narrow real actions such as restaurant bookings, and it is in request-based Early Access in the UK - free with Prime after Early Access, or £19.99 a month without.

Which AI agents can UK users actually get?

Supervised agents, mostly: ChatGPT Work (limited free, expanded on Plus at £20 a month), Claude Cowork (paid Claude plans) and Alexa+ Early Access. The always-on agents - Meta Muse, Gemini Task, OpenAI dots on personal plans and Claude Duet - are not available to UK consumers as of November 2026.

What can't AI agents do?

They struggle with vague goals, unfamiliar situations and websites full of logins and security checks, and they cannot reliably spot every trick hidden in the pages and emails they read. They need supervision on anything that matters.

Can AI agents be hacked?

They can be manipulated. In a prompt injection attack, hidden text in an email or web page instructs the agent to do something you did not ask for, and the NCSC says it is very possible this may never be totally mitigated. The published cases so far are researcher demonstrations rather than confirmed attacks on ordinary consumers.

What data do personal AI agents access?

Whatever you connect - email, calendar, files, signed-in websites - plus what they store in memory. Permission screens are usually all-or-nothing - the app decides what to ask for - so check what it requests before you grant, remove connections you stop using and check the training setting: some products use your data for training by default.

How much do AI agents cost?

The supervised agents UK users can try come with the AI plans: ChatGPT Plus is £20 a month and Claude Pro is £15 a month billed annually (£18 monthly) in the UK. The always-on agents mostly are not sold here; where they are live, Meta Muse has a free tier and OpenAI's dots come with its Pro plans - but not in the UK.

Are personal AI agents safe to use?

They can be used sensibly, but they are not risk-free. Start with low-stakes tasks, give the minimum access, keep approvals on and never share a bank login. The risk rises sharply once an agent can reach your email or spend money without asking.