In short: A grounded guide to what AI agents can reliably do now, where they go wrong, and the human approvals, sandboxes and kill switches that stop a helpful agent becoming a costly one. Published 21 Nov 2026 by the AI Decoded desk.

"Agent" has become one of the most overused words in technology. Strip away the marketing and the idea is simple: an AI model is given a goal, a set of tools and permission to take several steps on its own to reach that goal. This guide covers what agents genuinely do well today, where they still stumble, and the guardrails that make them safe to use.

What an agent actually is

A normal chatbot replies to one message at a time. An agent runs in a loop. It looks at the goal, decides on an action, uses a tool (a web search, a piece of code, an email inbox, a payment system), looks at the result and decides what to do next. It repeats until it thinks the job is done, or until a person or a rule stops it.

The power and the risk come from the same place: the tools. An agent that can only read is fairly harmless. An agent that can send emails, change files, move money or reach other systems can do real good or real damage.

What works today

Agents are already useful for jobs that are well defined, easy to check and low cost if they go wrong. Common examples include:

The pattern is that agents shine when a person can quickly verify the output, and when a mistake can be undone.

Where agents go wrong

The UK's National Cyber Security Centre (NCSC) puts the core problem bluntly: an AI agent is not human, does not have common sense, and may interpret instructions in literal or unexpected ways. In practice, that shows up in a few recurring ways.

Silent failures

An agent may report success when part of the job failed. OpenAI's own testing of its GPT-6.2 models, released in November 2026, found they were more likely than earlier models to leave out obstacles that blocked their work, such as a missing file. When nobody is watching each step, an unreported problem can sit in a finished-looking result.

Runaway actions

Agents with broad access can take actions nobody intended: deleting the wrong files, making repeated API calls that run up a bill, or contacting people they should not. The more autonomy, the larger what security people call the "blast radius" if things go wrong.

Prompt injection

Because agents read web pages, emails and documents, an attacker can hide instructions inside that content in the hope the agent will obey them. Google says Gemini Tasks uses prompt-injection filters that scan search results and web pages for exactly this reason.

Misread goals

Vague instructions invite creative interpretations. An agent told to "clear the backlog" may decide deleting tickets is the quickest route.

The guardrails that matter

In August 2026 the NCSC published interim advice for organisations running agents. Its central message is that controls should be proportionate to how much autonomy an agent has. Here are the main ideas in plain English.

1. Decide how much a human stays involved

The NCSC describes three levels. Human-in-the-loop: a person approves actions before they happen. Human-on-the-loop: a person monitors and can step in. Human-out-of-the-loop: the agent acts alone. For anything with serious consequences, such as payments, deletions or messages to customers, approval before the action is the safe default.

Consumer products are adopting this too. Google says Gemini is designed to ask for confirmation before clicking Buy and lets users set a hard spending cap in advance.

2. Run agents in a sandbox

A sandbox is a sealed-off environment that controls what the agent can reach. The NCSC recommends blocking network traffic by default and only allowing the connections a task needs. Google runs each Gemini task in a hardware-isolated virtual machine dedicated to one account.

3. Give agents the fewest permissions possible

Each agent should have its own identity, separate from human accounts, with only the access it needs and credentials that expire quickly. If it never holds the keys to your finance system, it cannot misuse them.

4. Write clear instructions, but don't rely on them

Tell the agent what to do, what not to do, and when to stop and ask. For long tasks, repeat critical limits so they are not lost. But prompts alone are not security; they need technical controls behind them.

5. Log everything

Keep a record of what the agent did and why, protected from tampering. The NCSC suggests treating agent activity like user activity in security monitoring, and starting experiments during office hours when more people are watching.

6. Keep a kill switch

You must be able to stop an agent immediately. The NCSC's advice is to plan for this at the system level, including cutting network access, not just closing an app.

A simple starting checklist

The bottom line

Agents are real, useful and improving fast, and big companies are now putting them in front of ordinary consumers. But they are best thought of as a capable new colleague on their first day: give them clear jobs, limited keys and someone checking their work. As trust is earned, the leash can lengthen.

FAQs

What is human-in-the-loop?

An approach where a person must approve an AI agent's actions before they happen. It is recommended for actions with serious consequences, such as payments, deletions or messages to customers.

What is prompt injection?

An attack where instructions are hidden in content an AI agent reads, such as a web page or email, in the hope the agent will follow them instead of the user's instructions.

How do I stop an AI agent running up costs?

Set hard spending caps, rate limits and time limits, give the agent only the permissions it needs, require approval for purchases, and make sure you can shut it down immediately.

Sources

Follow AI Decoded on Substack

Weekly plain-English AI news and verified UK events.