Different Tools for Different Jobs — Agent vs Chat (ft. Raya)

Short version: I spent half a day trying to get an AI agent to produce one clean PDF. Then I switched to a plain chat model and had the file in two minutes. This is not a story about a bad agent — it’s a story about picking the wrong tool for the job. Raya breaks it down in the video, and I’m writing it out here because I keep making the same mistake.

The PDF that ate half my day

The job was stupidly simple on paper: take a draft, clean it up, and hand me a finished PDF. No research. No multi-step workflow. No “go figure out what I actually want.”

I reached for an agent anyway, because agents are the cool new thing and everything in my feeds said that’s where AI is heading.

Half a day later, I had watched the agent spin up tools, open files, “plan,” reformat things I didn’t ask it to reformat, and hand me a document that was somehow worse than the draft I started with. Same trap I’ve written about before — the demo is always better than the first real session.

Then, out of pure frustration, I opened a plain chat window and typed the same request. Two minutes: clean PDF, exactly the structure I wanted, done.

Neither tool failed me. I failed at choosing.

Why the agent struggled

Not because agents are bad. They’re genuinely impressive — when the job is shaped like an agent job.

The agent I used built a plan, started executing, and kept going down a path. That’s its superpower… and its weakness. It wants to iterate, to take initiative, to handle things “end to end.” A single deliverable with clear formatting requirements doesn’t need any of that — every extra step is a new chance to add its own opinion, and every opinion it adds costs me time.

This is the exact thing Raya points out in the video: neither approach is bad — they solve different problems. An agent is built for the long game, and it expects to be told the shape of the problem. A chat model just answers.

The rule I’m stealing from this

Three questions before I pick a tool now:

1. Do I know exactly what the output should look like? If the answer is yes — I can describe the finished thing in one message — that’s chat territory. I’m not going to watch an agent rediscover what I already know.

2. Does the job involve steps I don’t want to babysit? Research that spans a dozen sources, updating files across a project, a workflow that has to keep going while I do something else? That’s an agent job. That’s where the half-day pays off.

3. Can I afford the iteration tax? Agents try things. Sometimes that’s exactly what I need — sometimes it’s five extra hours of formatting that nobody asked for. If I’m on a deadline, chat wins every time. If I’m building something that needs to keep working after I walk away, agent wins.

The honest bottom line

The AI world keeps insisting there’s a hierarchy: agents are the upgrade, chat is the older model. Real work isn’t ranked like that. For a single clean deliverable, the “outdated” chat model beat the shiny agent by about 23 hours and 58 minutes.

The tool that ships the work is the right tool. Grab the agent when the job has moving parts and a long horizon. Grab chat when you know exactly what you want and you want it before lunch.

Different tools, different jobs. Raya said it better in a minute and a half than I did in an article — but now you’ve got both.

Where this fits

This is the same take I had testing Claude Cowork: agents that act like teammates are worth real money when you define the work, and they’re a tax when you don’t. The tool has to fit the shape of the job — not the shape of the hype cycle.