Agents Are Here. Now What?

Every AI conversation I have been part of for the past several months has eventually landed on agents. It does not matter where it starts. Someone is describing a workflow, someone is building a tool, someone just finished a newsletter. Eventually: and then you could just build an agent to do that. The word has achieved a particular kind of ubiquity where everyone is using it and almost no one means exactly the same thing.

Let me try to be specific, because I think the vagueness is actually causing real problems in how people approach this.

An AI agent is not a chatbot with more steps. The distinction that actually matters for how you use them is about who owns the sequence of actions. A chatbot responds to you. An agent acts on your behalf. You give it a goal, it decomposes that goal into steps, executes them, evaluates the results, and adapts. You may not re-enter the loop until it finishes or gets stuck somewhere interesting.

I have been experimenting with agentic workflows since the start of 2025. My honest summary: some of it is genuinely transformative, some of it is genuinely terrible, and the gap between those two outcomes is almost entirely about how precisely the goal was specified before the agent started.

The things that work beautifully are the structured, bounded tasks. Research pipelines. Monitoring and alerting. Document processing. Content workflows with clear success criteria. In these cases, an agent can run largely unsupervised, produce reliable output, and give back hours that were going to mechanical work. That part is real and it is worth the setup cost.

The things that do not work yet are the open-ended, judgment-heavy tasks. Handle my inbox sounds like a dream. In practice, defining what handle means precisely enough for an agent to do it well, while making sure it does not send something you would regret, turns out to be a project in itself. Agents are very good at confidently going in the wrong direction. That is a specific kind of problem that is different from an agent simply failing.

What most people are underestimating is the new skill set that agents actually require. It is not prompting. It is systems design. You need to think in task graphs, not in questions. You need to define what happens when the agent encounters an ambiguous state: does it stop, ask, or make a call? You need evaluation criteria that let you trust the output without reviewing every step yourself. These are architecture skills, not chat skills, and most people building with agents right now do not have them yet. Including, sometimes, me.

The teams getting real value from agents in early 2025 started with one specific, well-understood process and built the agent around it. Not make our operations smarter. Here is the exact sequence of steps that happens every time we do X. Let us see which steps the agent can own. Specific. Bounded. Measurable. The question is never should we use agents. It is are we specific enough about what we want them to do.