What is an AI agent?

The definitional line

Three things distinguish an agent: a goal rather than an instruction, tool use to act on the world rather than only produce text, and autonomy over sequencing — the model chooses the next step based on what happened in the last one.

That third property is the whole thing. It's what makes agents able to handle situations nobody scripted, and it's what makes their failures hard to predict.

Where autonomy earns its cost

Agents are worth it when the space of situations is too large to enumerate. Debugging an unfamiliar codebase, researching across sources whose structure you don't know in advance, handling support requests whose resolution path varies — these are genuinely hard to express as workflows.

They're a poor trade when the sequence is known. If your process has six steps that never change, an agent adds latency, cost and non-determinism to something a workflow does reliably.

The cost of premature autonomy

The recurring failure pattern in early agent deployments is granting autonomy before reliability justifies it. Each step in a chain has an error rate, and those errors compound — a sequence of ten steps at 95% reliability succeeds around 60% of the time. Add the ability to take irreversible actions and the tail risk becomes the product's defining characteristic.

The teams that shipped successfully generally did the unglamorous thing: constrained the action space, put a human at the points of no return, and expanded autonomy as evidence accumulated. That's slower than a demo and considerably cheaper than an incident.

The product question underneath

Not "can the model do this" but "what happens when it does it wrong". If the answer is a wasted minute, ship it. If the answer is a deleted database or a sent email, the design problem is containment, not capability.