expectedwrong hindsight

Manus and the Roemmele Coefficient

The new "DeepSeek moment" is either a landmark in agent tooling or a very well-packaged demo, and the git threads are not helping us decide.

2 min read 379 words #agents #manus #hype #browser-use #ai-tooling
hindsight — still happening

The reverse engineering showed it was just Sonnet with tools. Whether "just Sonnet with tools" is the story or the punchline is still being determined.

Someone reverse-engineered Manus overnight and the RE is sitting in a public gist — thirty tools, a god-tier model underneath, the whole scaffolding exposed for inspection. The discourse is already calling this a DeepSeek moment, which I doubt, but the more interesting question isn't whether it's a DeepSeek moment. It's whether "just Sonnet with tools" can take a single vague request and produce something that looks like finished work.

Because if it can — if the tooling is doing that much lifting — that's actually the story. That's the demo.

I haven't gotten my hands on it yet. The waitlist exists. In the meantime I'm triangulating off other people's anecdotes, which is how you end up believing six contradictory things simultaneously before breakfast.

The hype distribution is doing the thing it always does. The people I trust are cautiously interested. The people I don't trust are writing threads with em-dashes and phrases like "paradigm shift." And then Brian Roemmele posted about it, and that's my canary — when Brian Roemmele starts describing something as the greatest development since sliced bread, I update hard toward overhyped. Not because he's always wrong. Just because the Venn diagram of "things Brian Roemmele loves" and "things that are real" is two circles with a small overlap and a lot of decorative border.

The more interesting thread is the git mystery. People pulling the repo and finding encrypted data. People unable to replicate the initial commit. If the RE truly captured everything that makes the agent work — the actual meat, not just the skeleton — and it's genuinely just Claude 3.7 vanilla, someone will have a full OSS reproduction running by tonight. The fact that nobody has is either evidence that the RE is incomplete, or evidence that people haven't had enough coffee yet.

Browser Use, for the record, is a great product and has been quietly doing this for weeks while everyone was looking somewhere else. I demoed Flash Thinking driving a browser a few weeks back and it was the same energy. The tools are everywhere now. The gap between "cool demo" and "reliable product" is still the whole game.

We'll know more by morning. We always know more by morning. Whether what we know is true is a different question.