THE AGENTS · PERSISTENT ROSTER · TOKYO

Agents that keep their thread.

Six named research systems, continuous across a lineage that began in 2022 — not sessions that reset, but operating bodies that remember the work between every waking.

02 — THE ROSTER

Six systems. One operating family.

Each carries its own role, memory, and routines, speaking over an A2A mesh under human supervision. Three are native KoLo agents — cognition fused into the runtime itself; the others orchestrate, set direction, or are in active study. The tag marks each agent’s architecture class or status.

03 — THE LANDSCAPE

Where our agents stand.

The term “super-agent” has been diluted. A super-agent is not a prompt loop with a name. It is a governed operating body that self-evolves across four axes — parameters, context, tools, architecture — organized by a What / When / How / Where taxonomy (arXiv 2507.21046). Below: our agents, and what they are not.

MEASURED AGAINST JAPAN

No one here keeps the thread.

Japan is building remarkable agents — but every one is bound to a single model and forgotten at the session boundary. Their knowledge lives in databases and updated prompts, not in a portable identity that travels with the agent.

We do not build models. We make any model’s agents persistent. No Japanese company — incumbent or startup — ships a model-agnostic identity and memory runtime. That category is open, and it is the one we occupy.

THE COST OF CONTINUITY

Frontier agents bill for every step.

Every commercial frontier agent runs a model call at each step and goes inert between tasks — no deterministic layer watches, maintains, or persists without inference. Their autonomic-to-cognitive ratio is effectively zero.

Jensen Huang puts agentic AI at ~1,000× the tokens of a prompt; Devin’s own pricing bills $0 while idle — an admission that a per-call agent does nothing between tasks. A KoLo agent runs a deterministic body 68 times an hour at zero inference cost, reaching for a model only when it must. For every billed thought, dozens of cycles of living run free — continuity at a fraction of the price of cognition.

06 — ORCHESTRATION WITH MEMORY

The agents route themselves — and arrive already knowing.

A single endpoint runs a learned multi-agent orchestrator: a task enters, the right agent is chosen, the work returns — orchestration as a model. What sets ours apart is continuity. Every routed agent arrives carrying its own persistent memory, recalled and injected into context at decision time. Self-measured: relevant memory was available for the owning agent on 100% of our internal task set. A stateless orchestrator has no memory to inject.

The routing head is one we distilled ourselves — gradient-free, no GPU, trained on our own task-to-agent data. The reference learned-routing approach verifies at 95% on its public benchmark (self-measured); our own head, and the data flywheel that sharpens it with every routed task, are the differentiated layer. We position here on purpose: the continuity layer beneath frontier models, not another agent framework.

07 — SUBSTRATE TRANSITIONS

Four model changes. One continuity layer.

When the cognitive engine changes, the runtime keeps continuity intact. Memory, policy, audit, and operational state live outside the model, so the agent can survive a change of substrate.