AI Agents: Useful Technology or Overhyped Automation?

"Agentic AI" has become one of those phrases that gets attached to almost anything with an API call in a loop, which makes it nearly useless as a category — some of what gets marketed as an agent genuinely changes what's possible, and some of it is a chatbot with extra steps. Separating the two matters if you're deciding whether to actually build a workflow around one.

Fact: an AI agent, in the meaningful sense, is a system that can plan a sequence of actions toward a goal, call tools or APIs to execute them, evaluate the results, and adjust its next step accordingly — largely without a human specifying each individual action. This is distinct from a single-turn AI response, and it's a real technical capability that's improved substantially as underlying models have gotten better at multi-step reasoning and tool use.

Where agents genuinely earn the hype

The workflows where agents work well share a common shape: a well-defined goal, tools that give clear, checkable feedback (did the test pass, did the file save, did the API call succeed), and a task that would otherwise require a human to do a lot of repetitive checking and adjusting. Coding agents that write code, run the test suite, read the failure, and fix it in a loop are the clearest example — the feedback loop is fast, objective, and cheap to run many times. Browser-automation agents completing a defined, repeatable task on a website are another.

Where it's a solution looking for a problem

Agents struggle badly, and are frequently oversold, in exactly the situations where feedback is ambiguous, slow, or expensive to get wrong — anything involving genuine judgement calls with no clear right answer, anything where a mistake is costly or hard to reverse, and anything where the tool the agent needs doesn't actually exist yet or returns unreliable results. A lot of agent demos look impressive on a curated example and fall apart on the messy, ambiguous version of the same task that shows up in real use.

Analysis: the useful question isn't "are agents good or bad," it's "does this specific task have the shape agents are actually good at." Tasks with fast, cheap, unambiguous feedback loops are strong candidates. Tasks that require weighing competing priorities, understanding organisational context that isn't written down anywhere, or accepting responsibility for a consequential decision are weak candidates — not because the underlying model isn't capable of reasoning about them, but because there's no reliable way for the system to know if it got it right, and the cost of being confidently wrong is high.

Opinion: a meaningful share of "agentic" product launches right now are agent-shaped wrappers around tasks that didn't need that level of autonomy — a simple, deterministic workflow dressed up as an agent because the term sells better than "automated script with a language model in it." That's not a criticism of agents as a technology, it's a criticism of how the label gets applied. The genuinely transformative uses tend to be quieter and more specific than the demos suggest: a coding agent that saves a developer real time every day, not a general-purpose autonomous employee.

Prediction, held loosely: the next few years likely bring real, durable value from agents in narrow, well-scoped domains with good feedback loops — software development, structured data processing, well-defined research tasks — well before "agentic AI" becomes trustworthy for open-ended, judgement-heavy work with real stakes. Expect the marketing language to keep outpacing the reliability for a while yet, and expect the actually useful agent deployments to look boring and specific rather than dramatic and general-purpose.