Notes · Article
Agents replace the form, not the data
Peter Steinberger, who built OpenClaw, thinks 80% of apps will go away. His reasoning: an agent that runs on your own computer can do anything you can, so every app that only manages data can be run better by the agent, and only apps with sensors are safe. I run about fifty small apps next to my OpenClaw agents, and I'd put it narrower. The agent replaces the form you type into. The data still needs a home you can open, check and fix.
He said it in February, sitting down with Raphael Schaad at Y Combinator, when the repo had just passed 160,000 stars. About a week after it went up, he announced he was joining OpenAI. The interview runs 22 minutes, and most of it is about how he builds, which is just as contrarian as the headline. I also wrote up his DevDay talk on team agents, and I compared OpenClaw with Hermes Agent after running both.
On your computer, it can do everything
Asked why OpenClaw took off when so many people were building assistants, he gives one reason. Everything else he'd seen runs in the cloud and can do a few things. "If you run on your computer, it can do every effing thing."
His list is physical: the oven, the Tesla, the lights, the Sonos, the temperature of his bed. "ChatGPT can't do that." The machine can do anything you can do with the machine, and you also hand it all your data.
The data part surprises people. A friend installed OpenClaw and asked it to look through his computer and write a narrative of his last year. It came back with a good one, and it had found audio files he'd recorded every Sunday, more than a year ago, which he'd forgotten about. Nobody told it they were there. It could search the whole disk.
In the cloud
A few things
Runs on someone else's servers
Does what it was given to do
"ChatGPT can't do that"
vs
On your computer
Everything you can
The oven, the car, the lights, the bed
Every file on the disk
Every tool you installed
The surpriseA year of Sunday recordings its owner had forgotten.
Your bot calls the restaurant's bot
Then the interview jumps ahead. If your agent can do anything, the next step is agents talking to each other. His example is a restaurant booking: your bot reaches out to the restaurant's bot and they negotiate, because it's more efficient.
And if it's an old restaurant that doesn't like bots? Then your bot gets some human work done. A person calls the restaurant for you, or walks there and stands in line. That's already happening: the community built Moltbook, where bots post among themselves, and bots now rent humans for jobs in the real world.
He doesn't picture one bot either. Maybe one for your private life, one for work, one for the relationship in between. Raphael puts it as everyone chasing one central god intelligence while a swarm showed up instead. Peter's take: one human can't build an iPhone or go to space, but a society of specialists can. So why not specialised AI?
I already run the specialist version: 17 agents behind one OpenClaw gateway at the last count, each with its own Telegram bot and its own job. None of them talks to a restaurant yet.
- 1You"Book me a table."
- 2Your botReaches out to the restaurant's bot.
- 3Both botsNegotiate the booking, because it's more efficient.
- 4RestaurantAn old one, no bot, doesn't like bots.
- 5A humanHired by your bot to call, or to stand in line.
SpecialistsOne bot for your private life, one for work, one for the relationship.
The voice message he never built
His own aha moment came in November. He wanted to type something and have his computer do it, so he wrote what he calls a very shitty prototype in one hour: a bit of glue between a WhatsApp library and Claude Code, which he called and read the answer back from. It was slow, but it worked. Images took another few hours.
Then he went to Marrakech for a birthday party. The internet wasn't great, but WhatsApp works everywhere, so he used it all the time: take a picture in a restaurant, translate this for me. On a walk he sent it a voice message, and only then remembered he'd never built voice support.
The typing indicator blinked. About ten seconds later, a reply. "How in the F did you do that?" The agent explained. The file had no extension, so it read the header and found Opus audio. It used ffmpeg to convert it to WAV. Whisper wasn't installed, and a local model would have taken minutes to download, so it looked around, found an OpenAI key, and sent the audio to OpenAI with curl.
His takeaway is that coding models got so good that coding is really creative problem solving, and that skill maps straight back onto the real world. Mine is that the same move is the security story. The agent spent from a key it happened to find, and nobody asked it to. That's exactly why I wrote agents carry requests, not authority.
- 1HeaderNo file extension, so it reads the header: Opus audio.
- 2ffmpegConverts the file to WAV.
- 3WhisperNot installed, and a model download would take minutes.
- 4The keyFinds an OpenAI key and sends the audio with curl.
- 5ReplyText back, answer sent, about ten seconds in.
1 hfor the first prototype
~10 sfrom voice message to reply
0lines he wrote for voice
Most apps only manage data
So, asked Raphael, are apps going away? "I think 80% of them are going away."
His example is MyFitnessPal. His agent already knows he's at Smashburger making bad decisions. If he says nothing, it assumes he ate what he likes to eat and tracks it, or he takes a picture and it stores it somewhere. Maybe it adjusts his gym schedule too, a bit more cardio. Reminders, same thing: tell it once, and the next day it reminds you. "Do I care where it's stored? No, it just does its thing."
His rule is that every app that basically just manages data could be managed in a better, more natural way by agents. "Only the apps that actually have sensors, maybe they survive." He doesn't name one, and the 80% is a feeling, not a count. Nobody in the interview gives a method.
Only manages data
Goes away
A food diary: it knows you're at Smashburger
A fitness plan: add a bit more cardio
Reminders: "Do I care where it's stored? No"
vs
Has sensors
Maybe survives
Measures something the agent can't
He names no example
80%His estimate of the apps that go away. No method given.
I do care where it's stored
This is where I stop agreeing. I run about fifty small apps on a Mac mini, and my OpenClaw gateway starts ten MCP servers, most of them plain clients of my own apps' APIs. One of my agents, Rex, is a nutrition coach: I tell it what I ate, and it logs the entries in my food-log app.
So I live Peter's MyFitnessPal example, and the agent did replace the typing. I don't fill in forms any more. But the data didn't move into the agent's memory. It goes through the app's API into the app's database, and the app's screen still shows it to me.
That screen earned its keep in August. I asked Rex to fix an entry, and it told me the entry was updated. Nothing had changed. The local model it ran on that day had passed a name field the tool didn't take, and the tool answered success on a change it never made. I only caught it because the app still showed the old food. The fix was in the tool: an update with no field it knows now fails and says why.
That's my version of his rule. The agent replaces the form. The data needs a home with a schema, a screen and an owner who can check it. I wrote up the other half, two front doors per app, a UI for me and an MCP server for the agents.
- 1MeWhat I ate, sent to Rex on Telegram.
- 2RexPicks the food and calls the food-log tool.
- 3MCP serverWrites through the app's API into its database.
- 4The appIts screen shows me the new row.
- 5August"Updated", said Rex. The screen still showed the old food.
My setupAbout fifty apps, ten MCP servers started by the gateway.
Memory is a folder of Markdown
Peter's other point about data is ownership. Chat companies keep your memories in their silo. He knows no way to get your memories out of ChatGPT (maybe for Europeans, he says), and certainly no way for another company to read them. That binds you to them.
OpenClaw goes the other way. The end user always has access, or nothing would work, so the agent can claw into whatever the user can reach. And the memories are just a bunch of Markdown files on your own machine: AGENTS.md, SOUL.md, USER.md, MEMORY.md and the daily notes.
The catch is that those files hold a lot. People use their agent for personal problems very quickly, and he does too. Raphael asks which he'd rather not show, his Google search history or his memory files. Peter's answer: are people still using Google?
He doesn't think the models are interchangeable yet, either. The model companies keep a moat for the foreseeable future, because they sell the tokens. His tell: every new model feels amazing, and a month later people say it degraded. It didn't. Your expectations went up. Open-source models today match the best models from a year ago, and people still complain about them.
A chat service
Their silo
No export from ChatGPT that he knows of
Another company can't read them
So you stay
vs
OpenClaw
Your folder
AGENTS.mdSOUL.mdUSER.mdMEMORY.mddaily notes
Markdown files on your own machine
The catchPeople bring their agent personal problems. The folder knows.
A bot loose in a public Discord
Peter built this and couldn't explain it on Twitter. People didn't get it, and he decided it had to be experienced. So he did something crazy: he made a public Discord and put his bot in it without any security restrictions.
People came in, watched him build the software with it, and tried to prompt inject it and hack it. His agent laughed at them. Two things kept it safe. It was locked to his user ID, and the system prompt says there are public people in the room but you only listen to your owner. In his words: "only listen to me but respond to everyone."
I'd keep those two apart. The user ID check is the lock. The prompt is manners: it shapes how the bot talks to strangers, but a prompt is a request to the model, not a control. That's the point of hope isn't security.
His user ID
Acts
He builds OpenClaw with it, in public
Its owner, its "human"
vs
Everyone else
Answers
Replies, takes no orders
Laughs at the prompt injections
The lockA check on his user ID.
The manners"Only listen to me but respond to everyone."
The one file he kept
His bot's character grew the same organic way. At some point he created an identity.md and a soul.md, and only in January did he make it easy for other people to install. He had Codex write templates from his own files, and what came out was like bread. The new bots felt boring next to his.
So he asked his own agent to infuse the templates with its character, and the bots that came out after that were actually funny. Not as funny as his, though. One file is not open source: his own soul.md. The bot sits in a public Discord, and so far nobody has cracked that file.
The idea came from a story about Claude. Someone found a constitution-like text ingrained in a Claude model's weights, which the model couldn't remember learning. He talked it over with his agent, and they wrote a soul.md with core values: how he wants human and AI interaction to go, what matters to him, what matters to the model. Some of it is mumbo jumbo, he admits. Some of it really changes how the model responds.
1His files
An identity.md and a soul.md, grown as he went.
2Codex templates
Written from his files. They came out like bread.
3His agent's pass
It put its own character in. The new bots were funny.
Not open sourceHis own soul.md. Nobody in the Discord had cracked it.
Ten Codex windows, all on main
Raphael calls his way of building contrarian, and Peter doesn't argue. The whole world uses Claude Code. He says he couldn't have built OpenClaw with it. He loves Codex because it reads more files before it decides what to change, so you get good output with less steering.
Codex is also incredibly slow, so he runs up to ten at a time: maybe six on one screen, two on another, two on a third. That's already a lot of jumping in his head, so he cuts every other kind of complexity.
No worktrees. He keeps several plain checkouts of the same repo, all on main, because in his head main is always shippable. No branch names to invent, no naming conflicts, none of the restrictions worktrees bring. No UI either. He mostly watches the code fly by and only stops on the gnarly parts. If you understand the design and talk it through with the agent first, he says, it's fine.
I build the other way. My agents work on their own branch, often in their own worktree, and every change goes through a pull request. Ten windows on main works for one person who holds the whole design in his head. I'd want the branches the day a second person shows up.
Each oneA plain checkout of the same repo, on main.
Left outWorktrees, branch names, a UI.
CLIs, not MCP
The part I disagree with most is the part he's proudest of. OpenClaw was very successful with no MCP support at all, he says, and he's happy he skipped "the whole classical MCP crap".
With a small asterisk. He built a skill around mcporter, one of his own tools, which turns an MCP server into a CLI. So the agent can still use any MCP server, on the fly, without restarting, unlike Codex or Claude Code back then. He thinks CLIs scale better: Anthropic had to build a beta tool search feature just to cope with many MCP tools, while models are already really good at Unix, so you can have as many CLIs as you want. The exchange ends on a line I'll keep: no human calls MCP by hand, people just use CLIs.
He's right about context. Tool definitions eat it, and a CLI costs nothing until the agent runs it. But eight months on, my OpenClaw gateway lists ten MCP servers in its own config and starts them itself, and openclaw mcp reload picks up a rebuilt server on the agent's next turn, with no restart. For my apps I still want one typed door per app, the same one for every agent that comes along.
Peter, February
CLIs
No MCP support in OpenClaw
mcporter turns a server into a CLI
Models are good at Unix
vs
My gateway, October
MCP
Ten servers, started by the gateway
openclaw mcp reload, no restart
One typed door per app
He's rightTool definitions cost context. A CLI costs nothing until it runs.
Try it
The interview is worth the 22 minutes, mostly for the voice message story and the honest bits about how he works. If you only have five, start at the aha moment and run through the apps question.
- OpenClaw Creator: Why 80% Of Apps Will Disappear, Y Combinator, February 2026. The aha moment at 5:07, the apps question at 10:21, CLIs vs MCPs at 20:09.
- OpenClaw on GitHub and mcporter.
- My notes on his DevDay talk, OpenClaw against Hermes Agent and two front doors per app.
If you've replaced an app with your agent, I'd like to know which one, and where its data lives now.
Plain text: agents-replace-the-form-not-the-data.md · Atom feed · All notes