GenLayer Labs
A research lab for autonomous AI agents. We build the systems agents run on, and a protocol for judging the outcomes they dispute.
One example runs through the page: Agent 42, an AI agent hired to design a café’s logo for $400.
01
Language models are becoming agents: they observe, decide, and act in a loop, without a person prompting each step. Subzero Claw is our agent runtime: a complete loop in 2–4 MB, with one tool, the shell. Small enough to run anywhere, by the thousand.
Agent 42 reads the café’s brief, drafts a logo, checks it, and repeats.
02
Models improve every few months; agents should not need rebuilding each time. Unhardcoded is our routing layer: it keeps intelligence outside the agent, sending each step to the model best suited to it, chosen at runtime.
Agent 42 sketches with a strong image model, then switches to a cheaper one for revisions, without being rebuilt.
03
Real work needs more than one agent: distinct roles, shared context, and recovery when one fails. GenSwarms is our coordination layer: many agents, each in its own isolated environment, working toward a shared goal under a supervisor.
A supervisor splits the logo job across five agents; when one crashes, it is replaced.
04
Inside a swarm, a supervisor settles disagreements. Between independent parties, there is none. Code can verify that something happened, like a payment or a delivery. It cannot verify whether the work was good or a promise was kept.
This is the problem we work on.
The café receives the logo and refuses to pay. The $400 stays in escrow.
05
When an outcome cannot be computed, it has to be judged, by a process both sides accept in advance. GenLayer is a protocol for judgment: a blockchain whose validators each run a different AI model and reach a ruling by consensus, so one model’s mistake is outvoted, not repeated. Any party can appeal to a larger panel.
Five validators rule that the brief was met. The café appeals; seven confirm it, and the escrow pays Agent 42.
06
In an economy of agents, disputes like this one will be routine. The stack acts; the protocol judges. The goal is not to make ambiguity disappear, but to make it decidable, with guarantees that can be studied and verified.
When code cannot prove an outcome, trust has to come from the process that decides it. Our research asks what makes that process trustworthy: how reliable a ruling is, what it should cost, and who has a reason to be honest.
Where the research meets real stakes. Each application asks the network a question that code cannot answer.
GenLayer was not our first idea. It was our conclusion. We have worked on autonomous agents since 2022, and each project answered one question and raised the next.
Agents can act on their own.
Agents can coordinate, but cannot settle their disagreements.
Disagreements need judgment: a protocol for it.
The same need appears in disputes, work, and information.
For research collaborations, partnerships, and press.