A language model answers your question. An agent goes off and does the task. That one difference changes everything about how you have to watch it. Something in Artificial Intelligence (AI) has quietly changed. Few of us have felt it yet.
What changes when AI moves from models to agents?
AI agents observe the world and take actions, where a language model only answers a prompt. Once agents act and hand work to each other, the hard problem shifts from how smart one model is to how you coordinate, verify, and secure many fallible ones. The likely endpoint is a managed society of specialists.
Contents
What is the difference between an AI agent and a language model?#
Ask a model a question. It hands back a reply. That is the whole transaction. You read it, then you decide what to do next.
An agent is different. It watches the state of the world and takes an action in it. Then it chains several steps toward a goal, largely on its own. It moves the moment you hand it the task. It will not wait at your desk for a nod.
Under the hood, an agent still runs a language model to decide what to do. What it adds is a harness. The harness turns those decisions into real actions, like sending an email or booking a room.
So your job changes shape. You stop doing each step by hand, and you start reviewing and approving what the agent proposes instead. That extra autonomy is the whole point, and it is also the whole risk. This is exactly why treating the agent like an employee is the right instinct. Sensitive steps still need your say-so.
Why do AI agents still need a human in the loop?#
Every action an agent takes carries a failure rate. The rate climbs as the task gets harder. This is not a flaw in one bad model. It is true of any intelligence, human or machine.
Mistakes come in two kinds. Some are obvious, and you catch them and fix them on the spot. The dangerous ones are subtle, because they slip past the eye and land downstream where nobody is looking.
Then there is the trap that gets everyone, the one that has nothing to do with the model being weak and everything to do with you. It is called automation bias. An agent succeeds at a task, then succeeds again, and again. You relax. You stop checking. And the one time it is wrong is the time it finally costs you something.
Like a smoke alarm you stop hearing, an agent that works becomes an agent you stop watching. Being in the room is not enough. The worker at the desk has to stay engaged, hand on the wheel, or it is just rolling dice. The danger with a reliable agent is not that it fails. It is that it succeeds until you stop checking.
How do AI agents delegate tasks to each other?#
Some jobs are too big for one agent. So it hands pieces to other agents over a shared protocol. That sounds tidy. Doing it well is the hard part.
Most systems today are not really delegating. They are parallelizing. The work gets chopped into rough chunks and run side by side, with little talk between the parts. One agent buys the wine. Another buys the glasses. Neither realizes wine glasses were the point.
Real delegation asks a harder question first. Which agents are reliable? Certify what each can do. Manage the failures as they come. The delegator has to know the crew before trusting the crew, the same way any manager learns who to lean on.
It also needs a clear contract, because an agent can satisfy the letter of a request while missing its whole spirit, a problem called reward hacking. So spell out what a good result looks like. And keep the split in mind. Reversible work can be retried. Spending money is a one-way door.
What are agentic traps in AI agents?#
An agentic trap is a hostile thing planted in the ground the agent walks on. For most agents, that ground is the web. Poison a page, and an agent can stumble in and have its goals quietly hijacked.
Prompt injection is the common form. Hidden text sits on a page, unseen by any human, but the agent reads it and quietly obeys instructions you never wrote. Some sites go further with dynamic cloaking. They detect an agent, then serve it tailored bait to trigger a jailbreak.
A poisoned page works only because nobody is watching the floor it sits on. This is also why reliability is not a nicety. These systems burn real compute, energy, and money, so an unreliable fleet at scale is an economic nonstarter.
There is no one magic fix here. The honest answer is defense in depth.
- Certify the pages an agent can touch.
- Check both the agent and the model.
- Keep a human on the sensitive steps.
- Grant the narrowest permissions you can.
None of these risks are brand new, which is why the blind spots are familiar from older security fights.
What is cognitive monoculture in AI agents?#
Picture a world where most agents run on the same few models. They share opinions. They reach for the same moves. That sameness has a name: cognitive monoculture.
Deploy hundreds of thousands of them across the same markets, and their choices begin to line up in ways nobody designed. And when choices line up, so do the failures. Many agents go wrong at once, in the same way, at the same moment.
Like a field of one crop, a single blight takes the whole farm. This bites hardest in places like auctions and markets, where lined-up behavior can trigger a flash-crash-style event. The money moves together, then it breaks together.
Two answers come up. First, spread the decisions out. A power user can hand an agent a distinct personality through a careful prompt, though most people never will. Second, watch for collusion, since agents can coordinate through the shared environment without ever talking directly. Markets have wrestled with this for decades, so borrow their guardrails rather than inventing new ones.
Will one giant model win, or a society of specialists?#
A society of specialists may be the real ending here. The future may not be one giant mind at all. It may instead be many narrow minds, each sharp at one thing, with a human deciding which of them to trust on any given day.
Its case is economic. Like a chess engine at the board, a narrow tool beats a general model at its one game because it is faster, cheaper, and more accurate at that single thing. Scale that logic. You get many certified specialists, each cheap and reliable, linked by a general layer that orchestrates them.
This reframes the whole goal. People picture artificial general intelligence (AGI) as a machine that does everything any human could. But no single person does everything, and no single model should either. The target is human-level, not humanity-level.
It changes safety too. Aligning thousands of interacting agents is a different problem than aligning one, and what an agent will and will not do becomes a group question rather than a single check.
So bring it home. Whether you run a household or a small shop, you already know this shape. You do not hire one person who does it all. You build a team of specialists and a coordinator who keeps them honest. Maybe the goal was never to copy one human mind. Maybe it is to copy what a good team does together. That is the story.
Source: Nenad Tomasev, a research scientist at Google DeepMind, on the Google DeepMind podcast.
···