Manus Is the Thing Everyone Claimed the Last Thing Was

Every few months the timeline fills up with takes about how some new agent is The One — the thing that finally crosses the line from "impressive demo" into "actually does stuff." Operator, Deep Research, Magentic. Each one was announced, each one got the breathless thread treatment, each one turned out to be a more sophisticated autocomplete for a narrow strip of tasks.

Manus is different, and I'm annoyed to say that.

It's a general autonomous agent out of China — Monica — and the demos show it doing the kind of multi-step, multi-tool work that previous systems would stub their toes on halfway through and then apologize. Browse the web, write code, execute it, read the output, adjust, write a report, manage the files. Not in a curated demo environment. In the world, where things are broken and weird.

The invite queue is already a nightmare. People are trading codes like it's 2009 Gmail.

What's actually interesting isn't the capability gap — it's the implication. If this holds up outside the demos, it doesn't just make Operator look like a toy. It makes the entire framing of "AI assistant" look like the wrong frame. An assistant waits for instructions. This thing just... goes and does the thing.

Nobody's calling it that yet because saying "autonomous agent" out loud still sounds like a 2022 venture deck. But that's what it is.

Whether it runs into the same walls everything else has run into — context limits, error propagation, the compounding fragility of long task chains — we'll find out in about two weeks when the people who got early access start posting their failure screenshots.

Until then, the buzz is warranted. Annoyingly.