I worked at Dropbox in 2014, pre-IPO. My job was to make your files more collaborative.
We couldn't dogfood our products, because the entire company ran on Google Docs.
The internal narrative was: desktop software is over. Files are over. The future is in-browser. Dropbox acquired Hackpad and focused on developing Dropbox Paper.
And that narrative was right, for a decade. Google Docs, Figma, Notion - everything moved to the browser, and the browser won.
Today, agents are reversing that. @claudeai Code/Cowork and @openclaw “just work” when you can point them to a folder on your computer.
Files are SO back.
But as it stands, files are not enough. Agents need more - they demand more. Where we're going, we'll need a whole new type of software.
The filesystem thesis
There's a post going around that argues that you can model your entire company as a filesystem.
If you’re a law firm: model your company as a set of folders. Cases go in /cases, timesheets go in /billing, etc.
Then point your @openclaw agent to the folder, and you’re in business.
This works shockingly well, and scales for text-based workflows like code, legal, ops etc.
Now, compare this to an agent trying to work with a web app.
Even with MCP, agents struggle to edit things like Figma files or Notion docs. It's slow, token-heavy, and breaks constantly.
Short term, it’s looking like stuff that can be accessed by unix tools and the command line is going to win.
Here's where it breaks down.
Agents need to see things.
A lot of work can’t be modeled as markdown instructions.
Think: designs, video, 3D models, circuit boards etc. The physical world. The multitudes of retail products and CPGs.
An agent can't operate on a photoshop, figma or CAD file in a vacuum. It’s going to need to “render” and see the outputs of what it’s doing.
These files are in legacy, proprietary formats. They are token-inefficient and hard to understand.
What about computer use? Can’t agents look at screens the way we do, and use the keyboard/mouse to click and operate these programs?
Maybe, but it’ll always be slower, more fragile and more inefficient than dealing with native files. Agents are most successful when they have a good harness. IMO, computer use doesn’t get us there.
Agents also need to show things!
The last decade of SaaS was onto something.
Web apps won because the browser solved human-to-human collaboration with real-time editing, commenting, team accounts etc.
Today’s filesystems and agent abstractions work for a 1 human : 1 agent model.
They don't work well for multiple humans : multiple agents.
Imagine human operators and agents working together to manage a power plant. Do you think chat is the best way for them to work together? You probably want some kind of UI where they can map out the site, identify things that need attention etc.
Could the agent generate this on the fly? You probably want to have built this in advance, so that there’s a fixed “contract” of how the system works and it doesn't change every time.
In many situations, you probably want a pre-built interface.
The tools that win will work on two levels: files & UIs.
So, in the brave new agentic world, what tools will win?
Tools that combine the best of both worlds:
1/ An open and agent-friendly file formats
Something an agent can read into a context window, reason about, and write back.
2/ UIs that work for agents AND humans.
Agents need ways to visualize output and take multimodal actions without going through a UI built for a mouse and keyboard.
Humans need ways to review, steer, and collaborate with agents in real time.
My prediction is that these tools of the future will not be built by Microsoft, Salesforce and Adobe. To win, they’d have to throw out the file formats they’ve calcified over decades.
Files are back. But the file format wars are just getting started.
Thanks to @ben8128 @danlovesproofs and @ravisparikh for reading drafts of this.