Get in touch
製品The Blossom platform

Ryōkai, Eval,
and Agents.

Products and capabilities for putting AI agents to work inside operationally complex businesses: task routing to agents and agent swarms, pre-deployment evaluation and improvement, and long-horizon agents — informed by what the lab is finding.

両界 · 評価 · 行動Routing, evaluation, and action.
— Product principle
両界

Ryōkai

From models to agents and swarms
Routes
Tasks to agents and agent swarms
Constraints
Budget · deadline · policy

Evolved from Furiwake’s model router, Ryōkai routes tasks to agents and agent swarms. It selects agents or complementary teams within budget, deadline and policy constraints, using the same task specification to evaluate candidates and verify delivered work.

Visit Ryōkai ↗

Agent routing

Routes each task to agents and agent swarms.

Constrained selection

Selects agents or complementary teams within budget, deadline and policy constraints.

Same specification

One task specification evaluates candidates and verifies delivered work.

評価

Blossom Eval

Pre-deployment evaluation and improvement
Stage
Pre-deployment · CI/CD
Approach
Origami + meta-heuristics

Blossom Eval is designed for pre-deployment evaluation and improvement in CI/CD. For well-defined objectives, it combines Origami and meta-heuristics to support RSI. For open-ended tasks, similar methods aim to establish a performance lower bound. Annotation supports the process where needed.

Talk to us →

Pre-deployment, in CI/CD

Evaluation and improvement before deployment.

Well-defined objectives

Origami and meta-heuristics, combined to support RSI.

Open-ended tasks

Similar methods aim to establish a performance lower bound.

行動

Blossom Agents

Long-horizon work
Approval policies
Per-step · per-tool · per-cost
Tool sandboxing
Configured per project
Replayable traces
Recorded where enabled

Long-horizon agents for operationally complex work. Human approvals where they matter — high cost, high risk, irreversible — and autonomy where they don't. Traces of decisions and their inputs can be recorded for review, replay and audit.

What is sandboxed, traced and gated by approval is set for each project with the team that runs it. Designed for the work where the model running unattended is exactly what you want, and also exactly what you fear.

Request access →

Approval policies

Configurable thresholds for when humans must sign off.

Tool sandboxing

Tool calls can run in isolated, observable environments, scoped to each project.

Replayable traces

Where tracing is enabled, a decision can be investigated after the fact from its recorded trace.

Cost-aware planning

Designed to budget the work before doing it.

原則Principles

Distinct jobs,
shared principles.

Ryōkai routes each task to agents or complementary teams and verifies the delivered work against the same specification. Eval evaluates and improves before deployment. Agents carry out long-horizon work. What they share are principles: evaluate before you trust, decide under explicit budget and policy, and keep decisions reviewable by the people who have to defend them.

The platform is research-informed but not research-locked. We ship improvements that have cleared an internal evaluation bar — and we publish the studies so you can check our work.

See the research behind it →