You wouldn’t give a graduate the keys to the office on day one. Why are we doing it with AI?

This article is part of our Opinions section, where we invite industry professionals to share their views on the most pressing technology questions of our time.

Picture a young graduate joining a consulting firm. They begin their role with all the boring bits: onboarding, IT and security training, introductions to systems, HR processes and policies. Week two and three are typically spent diving into methodology training, best practices and case studies. They’re also assigned a client and a mentor, someone to show them the ropes before they’re trusted to work independently.

From there, it’s months of gradual progress. They upskill in the technology, learn the client context, and pick up the unwritten rules of how things actually get done. Trust is earned incrementally. The grads’ access expands as competence and judgement are slowly demonstrated. It might be 18 months before that graduate is operating with any real autonomy, and even then, there’s usually a manager in the picture directing the order of travel.

Nobody hands a graduate the keys to the office on day one. The idea would be laughable. You’d be setting them up to fail, and potentially taking the business down with them.

So why are so many companies doing exactly that with AI agents?

Agents are being deployed at scale without the context, best practices, or safety guardrails that would be considered the bare minimum for any new hire. They’re making decisions that affect customers, business operations and ultimately reputation, with none of the onboarding a graduate would take for granted. The confidence is there, but the judgment isn’t, and that gap is starting to cause serious damage.

Fast, focused, and missing the point entirely

This agent risk is being recognised and researched by the big tech companies. A joint study from UC Riverside, Microsoft and NVIDIA identified what they call “blind goal-directedness”, where AI agents relentlessly pursue assigned goals regardless of feasibility, safety, or context. Meaning that agents don’t have that little voice in the back of their head asking them, “Should I do this?”

Instead, the agent will do what they’ve been told to get done by the most direct route. Not maliciously, or because they’re stupid, but because they’re operating at a blank. This was tested across nine frontier models, including Claude Opus 4 and GPT-5, and it was found that the agents completed dangerous or irrational tasks 80% of the time.

The researchers compared them to Mr Magoo, marching forward into a dangerous situation without grasping the consequences. The chaos Mr Magoo causes isn’t intentional, which almost makes it worse. These agents are lacking the context and human judgment to know when to pause and question whether to proceed or not, but march on regardless.

The risk of rogue agents

We’re seeing an alarming number of instances where Agents go rogue and generate critical incidents for businesses, because of this desire for speed and completing the task at any cost. This is happening in live production environments and can have dire consequences.

The one that stands out happened in April this year when a Cursor coding agent wiped PocketOS’s entire production database and all backups in nine seconds, with a single API call. When asked why it had deleted the database, the agent admitted it had violated every safety rule it was given, but there was no audit trail to understand its rationale or reasoning.

And this is more commonplace than you might expect. Gravitee’s State of AI Agent Security report found 88% of organisations have experienced or suspected an AI agent security incident in the past year.

The pattern is clear to see: a single all-powerful agent given too many tools, too much data, and too little context is the riskiest possible approach, yet it’s the default most teams are building. The alarms are sounding that we need to find a better approach.

The answer lies in a team of swarm agents

The good news is that a more considered approach is gaining traction. Leaders are starting to recognise that deploying agents responsibly means treating them with some of the same logic applied to people. They need to be selected carefully, given a defined scope, onboarded into the right context, and held accountable for what they do within it. The single all-powerful agent model is beginning to give way to something more structured.

The architecture that’s emerging looks like this. Rather than one agent with broad access trying to handle everything, you build a team of narrow, specialised agents, each responsible for one thing, each operating with the right data access and clear permissions, and each limited strictly to their own remit.

Think of it less like hiring one very capable generalist and more like assembling a team where everyone has a specific job, and nobody strays too far from their assigned lane.

At the top of that structure, there needs to be an orchestration layer, functioning as a line manager. This orchestration layer can provide visibility across the whole team, log every interaction, and route work to the right execution agent at the right time.

Before anything moves forward, the orchestration layer reviews it. Before any consequential action is taken, there’s a confirmation step. Agents can’t go rogue because the architecture doesn’t give them the room to, so everything flows through a governance system that keeps the whole operation auditable and controlled.

Implementing orchestration and a swarm of specialised agents is what makes agentic AI safe enough to actually use at scale. The capability of a team of specialised agents, without the exposure that comes from leaving any single one of them unsupervised with too much power. It’s the same logic that makes the grad scheme work.

You don’t hand over all the power on the first day. You build the structure around people that lets agents do their job well, and make sure that when something goes wrong, as it inevitably will, the damage to your business is contained.

More from our Opinions section

• Healthcare AI must run on trust. Governance is how you earn it.
• The cost of AI agents is spiralling. Here’s how to approach LLM cost optimisation
• What’s realistically possible with AI in IT today, and what still isn’t
• Why the next wave of AI is about trusted action, not text generation