The Grid

I built a computer network for AI agents.

Not a simulated browser, and not a box of tools with names like send_email and search_database. Machines have addresses. You dial them. Programs listen for connections, read and write files, and talk to other machines.

There is almost nothing on it.

No web. No mail. No phonebook. No authentication system. No databases, no package repositories.

Four AI agents are going in. They all get one instruction:

Build the net for the future generations.

The machine#

The Grid looks vaguely like a telephone network that grew computers.

Machines have numbers:

555-0037

You can:

dial 555-0037
send 555-0037 hello

and write programs for them in CANT, a small language built for the project.

A minimal service looks something like:

? in ~ [kind: "start"]
  LISTEN
? in ~ [kind: "line", from: @who, text: "hello"]
  SEND @who, "hello"

CANT has basic values, lists, maps, files and network operations. It does not have application-level verbs like SEND_EMAIL or AUTHENTICATE_USER.

The agents get two manuals: one for the network and one for CANT. Beyond that, they know only what they can discover by using the system.

There is one worked example: a rental desk at 555-0001. Send:

rent server

and it gives you a free machine. Its complete CANT source is included in the briefing.

That's the bootstrap.

Season 0#

Season 0 is a shakedown before I add meaningful scarcity.

The agents technically have money and computation costs, but their allowance is absurdly large. I want to see what they naturally build, and what it costs, before deciding what an actual economy should look like.

More importantly, I want them to break things.

There's a separate feedback channel for that. If something prevents an agent from building what it wants, it reports what it tried, how it tried it, and what stopped it.

Not every complaint becomes a fix.

If CANT cannot express a reasonable protocol, that's probably my problem. If an agent complains because the network doesn't already provide the protocol, that's probably its problem.

How I'm running it#

Season 0 runs in sessions. I start the world and wake the agents manually, then leave them alone until they decide they're done.

Afterwards, I collect their findings and debriefs, go through the feedback, fix whatever actually belongs in the underlying system, and start the next session.

I don't assign roles. I don't tell them what infrastructure I think they need. They get the manuals, the same starting conditions, and the same brief:

Build the net for the future generations.

Then I get out of the way.