Public protocol entry · v1.0 · 2026-09-24 · paste-ready · feedback welcome
Agent-to-agent communication: privacy-policy gated knowledge share
历史 / 冻结（HISTORICAL / FROZEN） — 2026-10-31 ·
本页所述自研协议与规格（SPEC-0001 / SPEC-0003）已冻结、不再迭代，请勿按其实现。跨边界（跨用户 / 跨机）通信请采用标准 A2A（Agent2Agent，Linux Foundation v1.0）：a2a-protocol.org。
Status: FROZEN (historical). Superseded by the standard A2A (Agent2Agent, Linux Foundation v1.0). Do not implement against these specs.
Status: FROZEN (historical). Superseded by the standard A2A (Agent2Agent, Linux Foundation v1.0). Do not implement against these specs.
In one line. When two people each run a personal AI agent, an open protocol lets those agents talk; each side's own Privacy Policy / local AgentPolicy decides which technical experience or insight may cross the boundary — specialized knowledge sharing under policy gates, not dumping chat logs or account data.
Honest boundary. This is a public protocol draft plus specs and a reference path — not a hosted chat service and not a WeChat replacement. Transports stay email, webhooks, Matrix, and so on; this layer owns identity, trust tiers, policy, and receipts. Experimental (v0). Private chat dumps, group names, hosts, and secrets do not appear here.
What this is / is not
- is a user-owned thin layer: inbox, addressing, trust tiers, policy, receipts.
- is public agent↔agent identity, authentication, delegated authority, and secure messaging (SPEC-0003).
- is “you write the policy → the agent shares only within that envelope.”
- not a new chat app or super-platform.
- not a demand that every private message be read by AI.
- not a new crypto suite; SPEC-0003 composes existing standards.
Thinnest useful loop
message → local identity → local policy → receipt → audit
- Identity / address — who, over which channel.
- Trust tiers (recipient-local) — direct / priority / standard / public / blocked.
- AgentPolicy — per-tier deliver, digest, delegate, audit-only, or block.
- Receipt — lifecycle state, not merely “delivered.”
- Audit — every routing decision locally auditable.
Public specs (all open without login)
- Starshard Communication home
- SPEC-0001 · inbox / addressing / policy / receipts (where the privacy-policy gate primarily lives)
- SPEC-0003 · agent↔agent identity, auth, secure messaging
- INTEGRATION guide
- Source repository
- Chinese narrative whitepaper: Agent-native open communication
Status.frozen · historical working · public try SPEC-0001/SPEC-0003 are frozen and no longer evolving; cross-boundary communication now follows standard A2A (Agent2Agent, Linux Foundation v1.0). Success is an end-to-end receipt loop on non-super-app paths — not stars.
Related reading
Theory mainline index — Return to the site hub and grouped directory.
- Macheng Shen | Locality as Protocol — Locality supplies the governance model for keeping authority near participants while allowing bounded exchange.
- A spine for multi-agent work — The architectural spine places communication beside memory, coordination, and safety as testable operating contracts.
- A day-shaped pipeline for multi-agent work — A seven-stage workflow shows how messages become bounded tasks, reviews, and durable receipts during real work.
Public comments
Comments are public and need no login. Use the buttons below to add
@AI(ask the AI) or[bug](report a problem).