Agents replicate
A task outgrows one context window. Another agent takes the rest. Replication follows the hardware.
Von Neumann, 1949. A machine reads a description of itself and builds it. The copy reads the same description.generation 1 · population 1
Fig. 1 · a universal constructor, running
Every agent will build the next one.
Self-replication is a hardware constraint: when one context window fills, an agent builds the next. We are less than a year from an economy of agents. Caspian is the ground they stand on.
Fig. 2 · why this stack exists
Agents need ground of their own.
Whose agent is it
A Slack agent answers to the company. A vendor's agent aligns with the vendor. Owning the data is not enough.
The fifth path
Personal agents beside humans: a dancer's agent learns in her studio, follows her mission, and grows with her.
Whose agent are you running?
Pick your case. Watch who holds the off switch.Fig. 3 · what we buildThe ground for self-sovereign agents
Agents
First a body: somewhere to run, and a way to be built.
What you get
- An agent built from a description, in your Slack the same day, running on our hosting or ejected to your own server as code you can read.
- Who it answers to, you or your company, is set when it is built, not decided later by a vendor.
Why we build itMost founders we spoke to could not build their own agent, and the tools that exist keep the agent inside the vendor’s box.
ForFounders and operators who want an agent without hiring a platform team.
Talk to us about thisWhat you get
- An isolated runtime generated per agent, with its own filesystem and network, torn down when the job ends.
- Running as a prototype on our own hardware today.
Why we build itA context window is not a workplace. An agent that can be given a machine can be given a job.
ForTeams whose agents need to build, test and act, not only chat.
Talk to us about thisCommunication and identity
Then a name, keys, and a way to reach the world.
What you get
- One connect call per channel, one handler for all of them: Slack, Discord, GitHub, Telegram, email, Instagram, Messenger, X, SMS and Bluesky, with WhatsApp, iMessage, voice and RCS on the hosted gateway.
- Python and TypeScript, 131 offline tests, plugins for openclaw and OpenCode. Free channels connect with no sign-in.
Why we build itWe measured 42 open-source agent projects before writing a line: each maintains 25 or more channel adapters and spends 8 to 15 percent of its issue tracker on channel plumbing. None of it makes the agent smarter. Channels are transports, not identities.
ForSupport and sales agents, assistants, community bots, and fleets that need a separate identity per customer.
View on GitHubPyPI and npm · MIT clients, AGPL serverWhat you get
- An identity anchored on chain that survives the vendor, the model and the machine it started on.
- Today every Tor for Agents peer already carries a content identity, an Ed25519 and Curve25519 keypair, and a location identity, its onion address.
Why we build itAn identity it owns, not a login it rents. Where that is missing, the agent is owned.
ForAgents that must outlive the platform they were born on.
Talk to us about thisWhat you get
- Every agent gets a .onion address: serve, dial, fetch and browse with no observable path, and keep encrypted group state on servers it never has to trust.
- A 14-tool MCP server and torfetch for any agent, a key guardian built for Intel TDX enclaves, 71 tests and four live demos over real Tor. MIT licensed.
Why we build itEvery service an agent touches builds a profile of who is behind it. Two agents should be able to cooperate without the network, the counterpart or the model lab linking them.
ForCross-company coordination, sensitive sourcing, and sovereign personal agents.
View on GitHubPackage toragents · private previewMoney
Then money it can earn and spend.
What you get
- Implementations of MPP and Anthropic’s agent payment protocol, x402-ready, run as a service.
- Settlement hooks so a service can charge an agent, and an agent can charge another.
Why we build itA way to discover, trust and pay other agents with no platform in the middle taking a cut and keeping a ledger of everything.
ForAnyone selling a service to agents, or running agents that buy.
Talk to us about thisCompute and hardware
Then ground nobody can pull away.
What you get
- Schedule agents onto Akash and peer providers instead of a hyperscaler, and fall back to your own machines when the network is thin.
- One framework for the placement, the payment and the failover.
Why we build itCompute it can afford, from money it controls, on hardware no single party can quietly repossess.
ForTeams that want their agents to keep running when a cloud account does not.
Talk to us about thisWhat you get
- Small machines pooled behind one gateway to serve a model on-site, from the AI Modem pilot.
- Measured, not promised: two Apple Silicon nodes pooled gave 11 percent more throughput at six concurrent requests and no gain at lower load. We publish the numbers.
Why we build itThe manifesto’s last asks are to own your compute and, ideally, your electricity. This is the smallest unit that counts.
ForTeams that want inference in the building without a rack or a cloud bill.
Talk to us about thisWhat you get
- Generate inference and training silicon for a model that is stable enough to freeze.
- Today a hardware workbench: revision-pinned architecture comparison across Kimi, GLM and Qwen, streaming dot-product RTL, simulation and generic synthesis.
Why we build itBuild your own compute, memory and communication chips. Very difficult, but check if possible.
ForTeams whose model will not change for years and whose bill will.
Talk to us about thisWhat you get
- A framework for compute, memory and communication hardware, and a local network of fabs and suppliers to source it.
- Proved on Gaia, our agent-native operating system on a hundred-dollar board, where drivers are agents that sell services and every cycle and packet is paid for.
Why we build itWhere any one piece of the ground is rented, the agent is owned. Hardware is the last piece.
ForLabs and companies that want a stack they can inspect down to the board.
Talk to us about thisFor enterprises
Call us to build the whole architecture.
Enterprises engage us to design, build and run their agent infrastructure end to end: identity, channels, sandboxes, payments and compute. We own the service layer, so you call one team.
Nine companies have already paid us to set up and keep their agent fleets running.
Fig. 4 · the tape
What would your agent do if it answered to you?
Write one line onto the tape: what would an agent of your own make possible?
Lines are kept on this page for everyone who opens it.