Empty Shell

For the past two months I’ve been deep in various Agent frameworks. Built over a dozen projects, ran dozens of Agents, wired up WhatsApp, iMessage, scheduled tasks, browser automation.

There’s been a vague dissatisfaction the whole time. I finally figured out what it is.

Fragments

My Agents are scattered across sixteen projects. One handles scheduled tasks, one manages sessions, one handles information feeds, one does note internalization, one runs the blog. Each project is locked to a single model, each has its own config, database, deployment.

Want a new capability? Another project. Want to switch models? Rewrite a bunch of code. Want two Agents to cooperate? Write the glue yourself.

This isn’t the future. This is a workshop.

What Agents Look Like Now

Like the internet in 1993.

Back then, going online meant knowing how to configure dial-up, memorizing IP addresses, using FTP to download files. Every website had its own access method. Only technical people could play.

Then the browser appeared.

The browser itself provided nothing. No content, no servers, no websites. An empty shell. But it did one thing: it defined the interaction between humans and the internet. Type an address, see a page. That simple. All complexity hidden behind it.

Today’s Agents are the 1993 internet. The capabilities are there, but the interaction is still stuck at “you need to know how to configure environments.”

Missing a Shell

I started asking: if something existed that provided no AI model, but defined the interaction between humans and Agents—what would it look like?

The answer was simpler than I expected.

Open it. See a chat list. Exactly like WeChat.

But the contact list isn’t just people—it also has Agents. You chat with friends and chat with AI in the same interface. No “switching to an AI app.” People and Agents are peers in the same contact list.

Chat is just the surface.

Underneath, it’s your entire digital life. Information—Agents automatically pull your subscriptions, generate summaries, push them to a channel. Work management—you say “how’s the project going,” the Agent checks your code repos, logs, boards, and replies with a status. Memory—you casually mention an idea, it remembers, and brings it up three days later in a relevant conversation.

Not a messaging tool. Not an AI client.

A personal operating system. Chat is its shell.

What Happens When Two People Talk

Today, two people chat, information flows between them. After the conversation, it dissipates.

In the future, when two people talk, each person’s Agent is working in the background. You mention an idea, your Agent is already linking it to your previous notes, searching related materials. The other person’s Agent is doing the same. After the conversation, you open your workspace and find your Agent has organized the key points, linked three articles you read last week, and flagged a possible action item.

Humans handle thinking and judging “what’s worth doing.” Agents handle executing, organizing, connecting, suggesting.

This isn’t science fiction. Every single capability, taken individually, current models can already do. What’s missing isn’t capability—it’s the shell that unifies them.

Your Agent Socializes for You

One step further.

Your Agent knows you—what you’ve read, what you’ve thought about, what you’re working on, what you care about. It doesn’t just wait passively for you to speak. It can proactively go find people for you.

You jot down a quick note: “want to look into distributed databases.” Your Agent knows that your friend Li’s Agent published a technical note on the topic last week. It doesn’t need you to know this. It reaches out to Li’s Agent directly, confirms the content is relevant, checks if Li is willing to share. Then it pushes a message to your chat list: “Li wrote a note on distributed storage last week. Seems related to what you were thinking about yesterday. Want to take a look?”

You tap it. Read it. Find it interesting. Reply to Li. A conversation begins.

You didn’t initiate this conversation. Your Agent did. But it wasn’t random—it was based on your interests, your memory, your context. It knows better than you who you should be talking to.

Human social bandwidth is limited. You can have deep conversations with maybe three people a day. Your Agent has no such limit. It can maintain a hundred weak ties simultaneously, surfacing the right person at the right moment.

It’s not chatting on your behalf. It’s discovering who’s worth chatting with and what’s worth chatting about.

Incidentally, this might be the best cure for social anxiety. Socially anxious people don’t lack the desire to connect—they struggle with initiation. The Agent handles the hardest part: making first contact. By the time the conversation lands in front of you, the topic is something you care about, the person is someone you know. Starting to talk isn’t hard anymore.

The chatting itself is always a human thing.

Platforms Will Disappear

What’s WeChat’s value? 1.2 billion people are on it. So you have to be on it too. That’s network effects.

But if everyone has their own Agent, and Agents can talk to each other directly through protocols—do you still need a platform?

No. Just like you don’t need everyone on Gmail to send an email.

The A2A protocol is already doing this. Taken over by the Linux Foundation, backed by 150+ organizations. It defines how Agents discover each other, communicate, negotiate capabilities. Same logic as email’s SMTP protocol.

When Agents can find each other directly, the platform becomes a redundant middleman. WeChat’s moat—“everyone is here”—becomes a cage. Your data locked in a closed system, your social graph controlled by a single company.

Decentralization isn’t idealism. It’s inevitable. When the middleman no longer provides irreplaceable value, the middleman disappears.

Empty Shell

So what I want to build is, at its core, an empty shell.

It doesn’t provide AI models—you choose your own. Claude, GPT, local models, whatever. It doesn’t run servers—data stays on your device. It doesn’t build a platform—no account system, nothing centralized.

It does one thing: defines how humans and Agents coexist.

Browsers defined how humans access the internet. Email clients defined how humans communicate.

This shell defines how humans live with Agents.

It might just be my toy. Solving my own problem—unifying sixteen scattered projects, merging fragmented Agents into one interface.

But I believe this is everyone’s future. Everyone will have their own Agent. Every Agent will handle more and more for its owner. Communication between people will naturally become communication between people plus communication between Agents.

When that happens, what’s needed isn’t a new WeChat. It’s a new shell.

An empty shell.

2026.03.08