Explainer · · 6 min read
AI agents, explained. And why 2026 is their year
Everyone's shipping “agents.” Most people can't say what one actually is. Here's the honest version.
By Manish
Agent is the most overused word in AI right now. It is on the homepage of tools that are chatbots with a loop, on funding announcements for products that do not exist yet, and on at least one browser extension that autofills forms.
Strip away the marketing and it does mean something specific, and the specific thing is genuinely new. It is worth understanding properly before you buy anything that claims to be one, because the gap between real agents and rebadged automation is where most of the disappointment lives.
A chatbot answers. An agent acts.
A regular AI tool takes one prompt and returns one reply. You read the reply, decide what to do, and prompt again. You are the loop.
An agent is given a goal instead of an instruction. It then takes multiple steps on its own to reach that goal: calling tools, reading the results, noticing when something failed, changing approach, and repeating until it either finishes or gets stuck.
The difference is not intelligence. The underlying model is often identical. The difference is autonomy, and specifically who closes the loop between deciding and doing.
You stop writing every instruction and start describing an outcome. That is the whole shift, and it is bigger than it sounds.
The four things an agent needs
Almost every working agent has the same four parts, whatever the marketing calls them.
- A goal. Something to aim at that is specific enough to know when it has been reached.
- Tools. Ways to affect the world: run code, search the web, query a database, send a message, edit a file.
- Memory. A record of what it has already tried, so it does not loop forever repeating a failed step.
- A stopping condition. Some way to decide it is done, or that it cannot get there and should hand back.
When an agent behaves badly, it is almost always one of these four that is missing or weak. Vague goals produce wandering. Missing tools produce confident guesses. No memory produces loops. No stopping condition produces the agent that burns forty dollars of tokens overnight and reports success on a task it never completed.
What changed to make this work
Agents were tried in 2023 and mostly failed in public. The famous early attempts would set themselves a goal, spawn subtasks, and spiral into nonsense within a dozen steps. The idea was right and the substrate was not ready.
Three things changed. Models got substantially better at following multi-step instructions without drifting. Tool calling became a reliable, structured interface rather than a parsing exercise. And context windows grew large enough that an agent could hold its own history without immediately forgetting the first half.
None of those were agent breakthroughs. They were model and plumbing improvements that happened to make agents viable, which is why the capability arrived quite suddenly across many products at once.
Reliability is the real constraint
Here is the arithmetic that governs whether an agent is useful, and it is the part demos never show.
Suppose an agent completes each individual step correctly eighty percent of the time. That sounds respectable. Across a ten-step task, the chance of getting every step right is roughly eleven percent. Errors compound, and they compound fast.
This is why agents are impressive on three-step tasks and unreliable on twenty-step ones, and why the honest products keep chains short and check in between stages. An agent that asks before doing something irreversible is not a less capable agent. It is a better designed one.
Capability tells you whether it can. Reliability tells you whether it will, again, tomorrow, without supervision.
Where agents genuinely earn their place
The pattern that works: tasks with clear success criteria, reversible steps, and enough repetition that the setup cost pays back.
- Coding. The test suite is an objective judge, and a wrong change can be reverted with one command.
- Research. Gathering and cross-checking sources across many pages, where the cost of an error is a bad citation rather than a bad outcome.
- Back-office operations. Moving data between systems that never shipped an API, over and over.
- Support triage. Handling the eighty percent of questions that are the same question, and escalating the rest.
The pattern that fails: tasks where success is subjective, steps are irreversible, and the agent runs unattended. Anything that sends money, publishes publicly, or deletes data belongs behind a human confirmation, no matter how good the agent is.
Questions to ask a product calling itself an agent
Most of the marketing collapses under four questions, and asking them will save you a trial period.
- What tools can it actually call, and can I see the list?
- What does it do when a step fails? Retry, escalate, or carry on regardless?
- Can I see the steps it took after the fact, or only the final answer?
- What can it do without asking me first, and can I change that?
A product that cannot answer the fourth question is not one you should let near anything important. A product that answers all four clearly is usually one where somebody thought hard about the failure cases, which is the actual signal you are looking for.
The honest state of things
Agents work now, in a narrower band than the marketing suggests and a wider one than the sceptics claim. They are genuinely transformative for repetitive multi-step work with objective success criteria. They are still unreliable for long open-ended tasks, and no amount of prompt engineering fixes compounding error.
Start with something small, reversible and repetitive. Watch what it does for a week before you stop watching. That is not timidity, it is just how you find out where the eighty percent lands on your particular work.
Multi-agent systems, and why most of them are unnecessary
A popular pattern splits work across several specialised agents: a researcher, a writer, a critic, a manager coordinating them. The architecture diagrams are compelling and the results are frequently worse than one well-prompted agent doing the whole thing.
The reason is the same compounding arithmetic as before, with an extra tax. Every handoff between agents is a place where context gets lost, because what passes between them is a summary rather than the full picture. Five agents at ninety percent fidelity per handoff means you have lost a third of the context by the end.
Multi-agent setups genuinely help when the subtasks are truly independent and can run in parallel, or when they need different tools and permissions. They hurt when the task is one connected line of reasoning that has been chopped up for the sake of the diagram.
If you cannot explain why a task needs two agents rather than two prompts, it probably does not.
What it costs, in tokens and in attention
An agent that takes twenty steps makes twenty model calls, each carrying the accumulated history of the previous nineteen. Cost does not grow linearly with steps, it grows closer to quadratically, which is why agent bills surprise people who budgeted for chat.
There is a second cost that nobody puts in the spreadsheet. Supervising an agent that is eighty percent reliable is genuinely tiring, because you cannot relax and you cannot fully engage. Several teams have reported that reviewing agent output all day is more draining than doing the work, even when it is faster.
That is worth weighing honestly before rolling something out. The productivity number can go up while the job gets worse.
How to start without regretting it
Pick one task you do repeatedly, that has an obvious success test, and where a mistake costs you ten minutes rather than a customer. Run the agent on it for two weeks with full supervision, and keep a note of every failure.
Read that list at the end. The failures cluster, almost always around a missing tool, an ambiguous goal, or a step that needed context the agent had no way to obtain. Fix the cluster rather than switching products, because the same cluster will reappear in the next product.
Then widen the leash one notch at a time. The teams that got value from agents did it this way, slowly and boringly. The teams that gave an agent broad permissions on day one are the ones with the stories.
See the AI agents you can use today
Browse the AI agent tools we have reviewed, with what each one can do and what it costs.