Upcoming event · October 1, 2026 · Tokyo
日本語

SOOS エージェンティックAI OS セミナー（東京）

2026年10月1日（木）14:30–17:00（開場 14:00）· Incubation CANVAS TOKYO。AIエージェントのアカウンタビリティとガバナンス、製造業におけるSOOSの活用。
セミナーのご案内 →
 English

SOOS Agentic AI Operating System — Tokyo Seminar

Thursday, October 1, 2026 · 2:30–5:00 PM (doors 2:00 PM) · Incubation CANVAS TOKYO. Agent accountability and governance, and SOOS in manufacturing.
Event details →

Registration is on Peatix: peatix.com/event/5174746

Sovereign Object OS Project

19 open protocols that make agentic AI reliable enough to deploy at enterprise scale.

More control over what your agents do. More confidence in what they produce. More capability to deploy them where it counts.

Browse Drafts Gap Analysis SOOS Stack
Sovereign Objectis to agentic AIwhat the IP packetis to the internet
The universal primitive that makes ecosystem-scale deployment possible.

The IETF foundation is solid. Agentic AI moved faster than anyone expected.

SCITT, WIMSE, CAEP, SSF, VDAF are excellent standards for a world of services and APIs. Nobody designed them for autonomous agents acting on consequential objects: contracts, financial instructions, bookings, medical records. That is not a criticism. It is simply where the technology went next, faster than anyone anticipated. We wrote the narrowest possible specifications to fill those gaps, building on existing IETF work wherever it fits.

Intent Declaration

No common format exists for what an agent was authorized to do, by whom, with what scope, and for how long. Without verifiable scope, agents cannot safely be given wider authorization.
IDP →

Delegation Chains

Multi-agent pipelines pass authority without attenuation, revocation, or propagation guarantees. A compromised sub-agent has no standard containment path.
MAD →

Hard Behavioral Limits

No machine-readable standard exists for what an agent must never do, regardless of instruction. Jurisdiction-specific law as a package import, not a Cedar authoring problem.
CAP →

Human Escalation

When an agent reaches the edge of its authorization, there is no standard protocol for pausing and transferring control to a human principal.
HEM →

Full Session Audit

Existing audit standards do not cover the complete agent session lifecycle including partial completion, revocation state, recovery gates, and achievability. Debugging agent failures should not be reconstruction work.
GAR →

Better control means better software.

Today, every team building with AI agents solves the same problems from scratch — how to keep agents on task, how to stop them when something goes wrong, how to know what they actually did. SOOS solves those problems once, at the right layer, as open standards anyone can use.

An agent with a signed intent declaration can safely be given wider scope. An agent with progressive trust accumulation earns authorization based on its actual track record. An agent with a complete execution record is debuggable in minutes, not days. The engineering properties and the operational properties are the same properties.

Apache 2.0. No commercial version. Standards only work if anyone can implement them.

Build

The kernel behind the protocols.

soosproject.com is the home of the SOOS kernel and developer tooling. The protocol specifications live at soosproject.ai.

KernelSpec v3

The kernel specification: a decomposed spec set covering the Rust crate structure, Cedar policy engine and GEC abstraction layer.

Rust Implementation

Reference implementation of the SOOS protocols. Air-gapped deployment support. No runtime dependencies on commercial services.

TypeScript SDK

TypeScript reference code and bindings, MCP server integration, and npm packaging for agentic coding systems.

Get updates

Releases, events and major draft milestones.

No marketing. Only substantive updates.
✓ Noted — thank you.
Something went wrong. Please try again.

The SOOS draft suite is a human-AI collaboration. The protocol architecture, standards strategy, gap analysis, and governing design decisions are the work of the human authors. The drafting, iteration, and text production were done in close partnership with AI. We say this plainly because transparency about how this work was made is part of what the work argues for.