"Agent" has become the word of the year. It's used for everything, and for sometimes very different systems. The line with a simple workflow seems to have blurred.
Yet it's clear-cut, as long as you look for it in the right place. Not on the side of the system's intelligence, but in a question more useful for deciding: who controls the trajectory?
A label that sticks to everything
The word "agent" is everywhere. Vendors, the press and teams use it for very different realities. Over time, it no longer designates anything very precise.
This confusion is natural. Marketing has loaded the word with promises, and every player pulls it toward their own product. Our goal here is simple: to give you the criterion that really settles it, then stick to it.
The intuitive criterion, and its limit
First intuition: an agent would be an AI that "thinks", that evaluates, that decides. It's appealing. But this criterion mixes up two distinct things.
An AI can rate the quality of a result, or choose between two answers, without directing the process for all that. Judging a step isn't the same as steering the sequence. That's the whole nuance.
The real criterion: who controls the trajectory
Here's the dividing line. The distinction isn't ours: Anthropic draws a clear line between the two. A workflow orchestrates models and tools through predefined code paths. An agent, on the other hand, dynamically directs its own process.
To find out who's at the helm, three questions suffice. Who decides the next action? Who decides the number of attempts? Who decides when to stop? If the answer is "the code", you have a workflow. If it's "the model", an agent.
Engineer Simon Willison sums up the agent in one formula: a system that runs tools in a loop to reach a goal. The goal sets the stopping condition. The loop, meanwhile, is directed by the model.
The same incident, two responses
Take a concrete case. Your system enriches a client record by querying an external service. One day, that service stops responding. Two architectures react very differently.
First case, the workflow. The code has the fallback planned: if the main service fails, try the backup service, then check the result. A control loop runs on a fixed step. Everything was anticipated by the team.
Second case, the agent. No fallback is written in advance. The model notices the failure, then chooses for itself what to try: another source, a search, a rephrasing. It's the model that charts what comes next.
Same incident, same objective. The difference isn't the intelligence deployed. It's who decides the course of action. That's the criterion, made tangible.
Three cases that lend themselves to confusion
A few cases lend themselves to confusion. None of them alone is enough to make an agent.
- Self-evaluation alone. An AI that rates its own result stays within a fixed framework, as long as it doesn't choose what comes next.
- An evolving reference. Updating rules or data doesn't give the model control of the sequence.
- The "adaptive" system. Adjusting to an input changes the result, not necessarily the trajectory. The path can still be charted by the code.
The common thread: an AI can be present, useful, even sophisticated, without ever directing the process. AI presence and agency are two separate questions.
A dial, not a switch
The boundary isn't a wall. Between the fully charted workflow and the agent that decides everything, there's a gradient. Autonomy is tuned, not declared.
A system can be a workflow overall, with a single step left to the model's initiative. That's often the healthiest setting: you open up control where it's useful, and keep it closed everywhere else.
Opening up autonomy where it pays off
Moving from the first case to the second isn't a vocabulary choice. It's an architectural decision. And it comes at a cost.
When the model directs, outputs become less predictable. The path taken is harder to trace and explain. Validation and security demand more attention, because you're bounding a freer behavior.
None of this condemns the agent. It's a reminder of a common-sense rule, one that Anthropic also advocates: only open up autonomy where it visibly improves the result. Elsewhere, a well-scoped workflow costs less and is easier to keep under control.