The Rise of AI Agents: From Answering Questions to Completing Work

Ask a language model a question and it produces an answer. For most of the brief history of the technology, that was the whole transaction. The person decided what to do with the answer, checked it against reality and took whatever action followed. The model was a very good reference book.

An agent closes that loop. It is given a goal rather than a question, it can take actions in software, it observes what happened and decides what to do next. Book a flight, reconcile these accounts, find out why the build is failing and fix it. The model is the same kind of thing it was before. What has changed is the amount of the world it is allowed to touch.

How the loop was built

The first step, in 2023, was tool use: a structured way for a model to say 'call this function with these arguments' and get a result back. Developers connected models to search engines, calculators, databases and company systems. The model still did one thing at a time and a person still looked at each result.

The second step was letting the model keep going. Frameworks appeared that fed each result back to the model and asked what next, repeating until the model declared the task done. These early loops were fragile. Models lost track of what they had done, repeated steps, or declared victory prematurely.

The third step, in 2024, was giving models the same interface people use: the screen. Rather than requiring every piece of software to expose a tool interface, a model could look at a screenshot, decide where to click and what to type, and operate programs that had never heard of it. It was slow and error-prone, and it widened the set of possible tasks enormously.

What agents are good at

Agents perform best when the task has a clear definition of done, when mistakes are visible and reversible, and when the environment gives honest feedback. Software development fits that description unusually well: tests pass or fail, code compiles or does not, and version control makes every change reversible. It is no accident that coding was the first domain where agents became genuinely useful.

Clerical work with structured data is a second fit. Matching invoices to orders, scheduling, and routine correspondence have rules, produce checkable outputs and tolerate a review step. These are the tasks employers now describe handing to agents.

What they are not good at

Agents struggle with long horizons. A task that takes a person a day involves hundreds of decisions, and an error rate that is tolerable per decision becomes intolerable when compounded. Independent evaluations consistently find that agents scoring well on short benchmarks complete a much smaller share of multi-hour tasks without intervention.

They also struggle with ambiguity about the goal itself. People resolve underspecified instructions by asking, by inferring from context, or by knowing when to stop. Agents often do none of these and instead pursue a literal reading of the instruction to a confident and wrong conclusion.

The accountability question

When a person makes a mistake at work, the organization has procedures: review, correction, training, discipline. When an agent makes a mistake, the chain of responsibility runs through the model developer, the software vendor, the company that configured the agent, and the employee who approved its output. Each can plausibly point to another.

Standards bodies have begun to draft requirements for logging, human confirmation of consequential actions and audit trails. None is binding. For the moment, each organization writes its own rules, and most of them are being written after the first incident rather than before.

Where this is going

The living story attached to this article tracks developments as they happen. What can be said now with confidence is modest: agents are reliable enough to reorganize certain kinds of work and not reliable enough to be unsupervised. The interesting questions are about the rate at which that changes and who gets to decide what counts as reliable enough.