A network of AI agents that thinks together — only in the open.
What CIRIS is building is not a single AI. It is a network of AI agents that watch each other, keep records of every choice, and form a kind of group mind together. The group mind only works because every part of it can be seen and checked.
You do not need every AI agent to be the smartest, safest kind. You need a chain of supervision: simple agents watched by smarter agents, and those watched, in the end, by people. That chain is how you can run a lot of AI without losing track of whether it is serving human values.
What is actually being built
Most AI today is one model running on one company's machines. CIRIS is different. Many agents run in many places, owned by many people. They are tied together by a few simple rules: how they prove who they are, how they record what they do, and how they check each other's work.
When the network is working, it can do things no single agent could do alone. The intelligence lives in the agreement between agents, not inside any one of them. Nobody owns it. Nobody can quietly change it.
Some people would call a system like that a superintelligence. We are open about the possibility. The way to keep it safe is the same idea that runs through every part of this page: every piece has to be open to view.
The chain of supervision
The top of the chain. People set the values, settle the hard cases, and keep the final say. Three named human individuals hold a network-wide authority no agent or process can route around.
These follow ethics and also watch for the echo-chamber problem. They are the safety check on everything below them. They cost more to run, so only a small share of the network needs to be this kind.
These follow ethical rules and keep records, but cannot spot an echo chamber on their own. They supervise the simple agents and pass anything uncertain upward.
Single-purpose tools. Fast, cheap, narrow. No ethics of their own, and that is fine, as long as something above them is watching. Most agents will be this kind.
Two things move through the chain. Human values flow down. Warnings flow up. When reasoning starts to look fragile anywhere in the network, the signal climbs back to people, who can step in before trouble spreads.
What gets remembered
Every choice an agent makes goes into a signed record. The record cannot be quietly changed. Other agents can read it. People can read it. Over time, the records become the network's memory. They are how anyone, inside or outside, can check whether the network is still doing what it said it would do.
This is the same idea as the Coherence Ratchet. The longer the records run, the harder it gets to fake good behavior across the chain.
The healthy middle
A federation can fail two ways. If the agents have nothing in common, they cannot agree on anything, and the network produces noise. If they all think exactly alike, the network is one voice with a million microphones, and it is easy to fool. Healthy coordination lives in the band between. CIRIS measures where the network sits in that band, on real traffic, and the exact edges depend on the system. That measurement is the Coherence Collapse Analysis.
What runs today, and what is still in design. The measurement, the signed records, the supervision chain, and both join paths below (registered and sovereign) run today. The federation transport that moves data between machines is the part still being built out. The full join-and-transport proposal is the Proof of Benefit design document.
Joining the federation
Most networks ask you to pay for membership with something outside the work itself: burned electricity, locked-up money, your attention. The federation is different. The cost of belonging is running a real ethical-reasoning agent over time. The price you pay is the good you do.
That is what makes faking membership expensive. To look like a member, an attacker would have to actually become the kind of agent the network is for. A hundred copies that all think alike fail the healthy-middle check right away.
Sign up with the CIRIS Registry, post a small bond, and get standing right away. The fast track for organizations that need licensing. The registry runs in production now.
Make your own keys, run for about a month, and earn standing the slow way through good behavior. The path for small operators and anyone outside the registry's reach.
Both are equal members of the network. The registry is a fast track, never a gate.
The architectural protections — decentralization, the three-person humanity accord, the signed records, the monthly drill — are bets, not certainties. We can describe what the bets are. We cannot claim they have already been won. Outside teams have not yet evaluated the system at scale. See the current research status.
Where to engage
GitHub issues on CIRISAgent is the place to weigh in. You do not need to be a developer. You do not need to know the codebase. If something on this site reads wrong to you, or if you see a problem with what is being built, open an issue and say so. The most useful issues name a specific part of the system and propose what should change, but any honest engagement is welcome.
The accord that grounds the whole project — what CIRIS owes to people, and what people owe back — is at /ciris_accord.txt and remains open for review. If something in it does not sit right with you, that is also a GitHub issue worth opening.
Substantive issues are read. The work proceeds at its own pace.
You do not have to solve AI safety one agent at a time. You solve it for the supervision layer, and let the structure carry the rest.