Are chats still any use?

These days a good chunk of my work is done by agents and harnesses, and yet I still open a chat window several times a day. I stopped to look at why, which tool I send each thing to and where I've got it wrong. At the end there's a three-question rule that has worked for me.

You'll have heard the fashionable line: "if you're still using chat, you're falling behind". They say it in videos, threads and talks, almost always with an agent working wonders in the background while the presenter has a coffee. The message is clear: chat is for beginners and the serious stuff is elsewhere.

I confess I bought it too. It was early 2026, the first time I left Claude Code, which I'd been using since 2025, working alone on a server while I watched. I thought: "that's it, chat is a thing of the past". That same afternoon I opened a chat to ask it how to word a delicate email to a client. I laughed at myself, but the question stuck.

I come from where almost everyone comes from. I started with ChatGPT, switched to Claude (I explain why in Why I switched from ChatGPT to Claude) and today I use several assistants: mostly Claude, but also OpenAI, DeepSeek, GLM, Kimi and MiniMax. And for all of them, the way in was a chat window.

So the question in the title deserves an honest answer. The short one is yes, chats are still useful. The long one is that they're not useful for everything, and it took me a while to learn what they're not for.

Before going on, three definitions for beginners, because these words get used a lot and hardly ever explained.

An AI chat is what everyone knows: you type, it answers and you do the rest. If it suggests a text, you copy it. If it gives you a command for the computer, you run it. The model thinks and talks; the hands are yours.

An agent is something else: you give it a job and it carries it out using tools. It reads files, searches the internet, writes documents, runs commands. You say what you want and it decides the steps.

The harness is the program that surrounds the model and gives it hands, memory and limits. Claude Code, OpenClaw and DeepSeek Harness are harnesses. When you have one installed on your machines, the agent is no longer a one-off: it lives there, with access to your files and your servers.

And one word that runs through all three: the prompt. It's simply the message you write to the model, the request or instruction. Whatever you call it, if you've written something to an assistant, you've written a prompt.

Here's the core of it. There are four uses where chat still beats the rest for me, and not out of nostalgia.

The first is thinking out loud before deciding. A few weeks ago I was torn between setting up a new service on a VPS or on the internal network. I opened a chat, explained the case and spent twenty minutes going round the pros and cons. I didn't want anyone to do anything; I wanted someone to argue with.

The second, writing and editing texts. Emails to clients, proposals, the first draft of the odd blog article. Here the prompt is almost a conversation: I ask for a text, read it, say what I don't like, it rewrites it. I copy what comes out myself once I'm happy with it.

The third, quick questions. What an error I'm getting means, what an option in a program is called, what the difference is between two formats. Setting up an agent for that would be like taking the car to go to the kitchen.

The fourth, learning something new step by step. When I started with Unsloth to fine-tune models, I learnt it in a chat, question by question, stopping every time I didn't understand something. An agent would have done it for me without my learning a thing.

Why does chat win in these cases? Because it's a conversation, because I have full control and because nothing gets touched unless I copy it. And there's another, less obvious reason: in a chat, my mistakes are cheap. If the answer is bad, I throw it away and nothing has happened; no broken file, no email sent off on its own. As for which one I use, for thinking and writing, Claude almost always; for a second opinion, OpenAI or DeepSeek; and for quick questions, whichever one I happen to have open.

There are multi-step tasks I used to do myself by copying and pasting between the chat and my programs. Those now go to an agent.

An example from this week: going through a folder with twenty of a client's documents, pulling the dates and amounts out of each one and putting them in a table. Another: looking up information on three suppliers, comparing it and preparing me a summary with links.

Why not chat? Because I'd have to act as the go-between: open a document, paste it, copy the answer, move on to the next. The model would think and I'd be the messenger, which is exactly the job I want off my plate.

Mind you, an agent isn't a trusted employee from day one. It makes mistakes, makes things up and sometimes does more than you asked. I cover the precautions in The trouble with working with agents.

The difference between a one-off agent and an installed harness is the difference between calling a plumber and having one living in your house.

With Claude Code installed on my machines, I ask it for things like reviewing a server's logs, finding out why a service has gone down and proposing the fix. It works directly on my files, and the next morning it can pick up where it left off because it has its notes. With DeepSeek Harness I've spent ten days running similar tests, and I write about it in Ten days working with DeepSeek Harness.

What I gain is continuity, access and automation: I don't have to explain every time where everything is or paste the context in again. What I lose is peace of mind: more risk, because it has its hands inside my systems, and more oversight, because I have to check what it does. It isn't free, and it's worth knowing that before installing your first one. A poorly supervised harness can do damage in a minute that would have been impossible in a chat, simply because in the chat it had no hands.

So as not to rely on impressions, I noted down everything I did with assistants during a normal week, the second week of September. The rough count:

What surprised me when I did the count: by number, chat won by a landslide, forty-six out of sixty tasks. But in hours saved, the harness took the lion's share: I reckon about six hours that week, against one or two from chat. Neither is surplus; they do different things.

I've got it wrong in both directions, and neither one makes me look good.

Using chat for something that called for an agent: I spent a whole afternoon pasting forty emails one by one into a chat window so it would summarise each one, and copying the answers into a document. Four hours of copy and paste for something an agent later did in fifteen minutes. What I learnt: if I catch myself repeating the same action more than five times, I'm using the wrong tool.

Sending an agent something that called for conversation: I asked an agent to prepare a proposal for a client from my notes. It did it quickly and nicely formatted, but it had misunderstood what the client needed and the proposal went off in another direction. If I'd talked it through in a chat first, I'd have spotted the misunderstanding by the second question. What I learnt: when judgement is needed, you think first and act afterwards.

After all this, I ask myself three questions before opening anything:

1. Do I need to think, or do I need it done?
2. Will it be repeated?
3. Does it touch my files or my systems?

If it's thinking, chat. If it's doing something once, agent. If it's doing something often and inside my machines, harness.

It's not a law, it's a habit that saves me mistakes. And it has helped me answer the question in the title without drama: chat hasn't grown old, it has settled into its place.