I believe software should be fast, reliable and invisible. Think about Google Maps. I use it every day and I never think about the software. I just think about where I’m going. The routing, the traffic data, the satellite imagery. All of it disappears, I’m just driving. That’s my standard for any technology: the software should disappear into the task.
But do today’s AI Agents meet this standard? I’m using ‘agent’ loosely — any AI that performs tasks on my behalf, whether that’s a chatbot, a coding assistant, or something more autonomous.
Fast – yes, with a caveat
I recently had Claude review a draft legal document and it provided useful, specific feedback in a few minutes. Faster than I could have done it myself, and faster than a lawyer. By any measure, AI Agents can complete complex tasks and cut hours of work into minutes.
But as I ask Claude to take on more complex projects, I find myself waiting. I task switch. I get up to do a personal task. I browse my phone. The process is reminiscent of old-school debugging: edit, re-compile, wait, run, test and edit again. AI Agents are fast, but the waiting changes my engagement with the software.
That waiting tells you something else: you shift from doer to manager. While the agent works, you are not in the flow. You are supervising. For some, that feels like a productivity gain. For others, especially those who take pride in craft, it feels like a step removed from the work itself.
Reliable – less so
I find I need to check the work constantly. Agents don’t make grammatical errors, but the wording can be awkward. And mistakes are still common. Sometimes the agent genuinely believes it completed a task that it cannot do, like resizing an image, or it gives me a link that doesn’t work. And agents still hallucinate – make stuff up.
There are also cases of what I’d call confused brilliance. I asked Copilot to determine the location of a photo. It did an impressive job analyzing the image – identifying landmarks, vegetation, architecture – but it overlooked the GPS metadata entirely and got the answer wrong. Showing astonishing skill in one area, yet blind to the simplest clue.
The bigger frustration is unpredictability. I can adapt to a tool that fails in consistent, known ways. What’s harder to work around is the confident wrong answer, the task the agent is certain it completed but didn’t. That unpredictability is what makes it hard to trust.
Invisible – not yet
There’s a reason it’s called “Prompt Engineering.” That phrase alone tells you the problem: humans are being trained to talk to the machine, not the other way around.
Here’s a concrete example. Anthropic specifically recommends using XML tags over markdown or lists to get better results from Claude. I understand why: clear structure produces clearer output. But XML is not how I think. I think in narrative, in conversation, in lists at most. When I have to reshape my thinking to fit the tool’s preferred format, the software is no longer invisible. I’m serving it, not the other way around.
Invisibility arrives when the agent conforms to the human—not when the human conforms to the agent.
The path forward
Does this make me a detractor? Not at all. AI Agents have already become the next level of basic productivity tools. Not working competently with an agent is as limiting as not being able to use a word processor. It’s now essential for anyone working in tech – and will soon be for anyone working at a desk.
AI may still be at the WordPerfect stage – powerful but requiring complex instructions to get real work done. Let’s do better at making AI accessible.
AI Agents won’t replace human ingenuity—they’ll extend it. The real breakthrough will come when agents feel like true partners: fast enough to keep us in flow, reliable enough that we stop double-checking, and invisible enough that we stop translating.
Which of the three – fast, reliable, or invisible – matters most to you? And where do AI agents fall shortest?
Leave a Reply