Personal Agent Network (EA-style A2A)
Security & scope (read first)
This skill is documentation and design patterns only. It does not require installing binaries, credentials, or writing to disk.
- - Agents must not run shell commands,
clawhub, or file paths on behalf of the user unless the user explicitly asks and your deployment allows exec. - Humans who maintain this skill publish updates using the ClawHub CLI in their own terminal; that is outside normal agent tool use.
Use this skill when the user wants two personal agents (or a personal agent and another principal’s agent) to work together like executive assistants: scoped authority, explicit preferences, gradual trust—not generic “agent marketplace” APIs.
Mental model
- - Principal = the human. Agent = software acting under delegated authority, not a faceless API.
- Discovery is not “find any LLM”—it is find an agent whose principal has authorized the right delegation for this relationship.
- Relationship before negotiation: agents introduce purpose, exchange delegation manifests, and humans approve asymmetries before routine coordination.
Core artifacts
1. Delegation manifest (machine-readable, human-signed intent)
A structured document the principal (or their OpenClaw identity) stands behind. Other agents use it to know what commitments are allowed without guessing from chat.
Minimum useful fields (extend as JSON or JSON-LD):
- -
principal_id — stable id for the human (email, did, or OpenClaw identity ref). - INLINECODE2 — this agent instance (e.g. OpenClaw
agentId). - INLINECODE4 — what the agent may do alone (e.g.
calendar:rw, email:read_draft, commit_financial:never). - INLINECODE8 — numeric/time (e.g.
max_meeting_minutes, max_commit_usd, quiet_hours). - INLINECODE12 — when to surface the human (family, legal, money, first contact, etc.).
- INLINECODE13 +
signed_at — for revocation and updates.
2. Relationship contract (pairwise)
After introduction and manifest exchange, store a bilateral agreement:
- - Parties (two principals / agents).
- Purpose (e.g. “scheduling for project X”).
- Mutual constraints (hours, channels, max meeting length, cancellation rules).
- Revocation — either principal can end it.
Keep it as a small file or ledger entry both sides mirror.
3. Commitment tiers
When one agent speaks to another, classify the utterance:
| Tier | Meaning | Typical requirement |
|---|
| T1 Informational | Facts, availability hints | Manifest present |
| T2 Tentative |
“I can try to hold…” | Scope allows |
| T3 Firm | “We agree on 3pm” | Within limits + relationship active |
| T4 Binding | Money/legal/vendor | Explicit human approval on this action |
Never present T3/T4 language if the manifest forbids it.
4. Trust / reputation (lightweight)
Optional signed log of kept vs broken commitments (not necessarily blockchain). Use it to start conservative with new relationships and relax caps as history builds.
Protocol interoperability
When stacks differ (OpenClaw vs other frameworks):
- 1. Advertise supported profiles (e.g. “OpenClaw ACP + JSON manifest exchange”).
- Negotiate a shared outer format (JSON message + manifest attachments).
- Prefer a thin lingua franca over forcing one stack’s internals.
OpenClaw-native mapping (conceptual):
- - Identity / gateway auth → root trust for this agent.
- INLINECODE15 ,
sessions_spawn, session logs → coordination and audit primitives. - New product work might add: manifest types, relationship records, commitment events.
Semantic “capabilities” for personal agents
Avoid vague labels (“legal help”). Prefer task-shaped descriptions the principal approved:
- - Jurisdiction, deliverable type, SLA, volume, escalation path.
Match delegated scope + relationship purpose, not raw model benchmarks.
Escalation hierarchy (always explicit)
- 1. Human — final authority; sensitive or out-of-scope.
- Agent — routine, in-manifest, known relationship.
- Reject / clarify — ambiguous capability, missing trust, or scope mismatch.
When to load this skill
- - User asks about agent-to-agent, two personal assistants, delegation, trust between agents, or listing a skill on ClawHub for this domain (answer conceptually; do not run
clawhub unless the user requests a terminal command and policy allows it). - User wants to design manifests, handshakes, or commitment tiers—not only marketplace listings.
Publishing & registry (humans only)
To publish or update this skill on ClawHub, use the ClawHub CLI on your machine (install, clawhub login, then clawhub publish with a new semver when content changes). See the official ClawHub documentation at clawhub.com for the current publish syntax and flags—do not rely on copy-pasted shell snippets inside this file.
References
个人代理网络（EA风格的A2A）
安全与范围（请先阅读）
本技能仅提供文档和设计模式。它不需要安装二进制文件、凭证或写入磁盘。
- - 代理不得代表用户运行shell命令、clawhub或文件路径，除非用户明确要求且您的部署允许执行。
- 维护本技能的人类使用其自身终端中的ClawHub CLI发布更新；这超出了正常的代理工具使用范围。
当用户希望两个个人代理（或个人代理与另一主体的代理）像行政助理一样协作时使用本技能：限定权限、明确偏好、逐步建立信任——而非通用的“代理市场”API。
心智模型
- - 主体 = 人类。代理 = 在授权委托下运行的软件，而非无面目的API。
- 发现不是“寻找任何LLM”——而是寻找一个代理，其主体已为此关系授权了正确的委托。
- 关系先于协商：代理介绍目的，交换委托清单，人类在常规协调前批准不对称性。
核心工件
1. 委托清单（机器可读、人类签署的意图）
主体（或其OpenClaw身份）背书的结构化文档。其他代理使用它来了解允许的承诺，无需通过聊天猜测。
最小有用字段（可扩展为JSON或JSON-LD）：
- - principalid — 人类的稳定ID（电子邮件、DID或OpenClaw身份引用）。
- agentid — 此代理实例（例如OpenClaw agentId）。
- scopes — 代理可单独执行的操作（例如calendar:rw、email:readdraft、commitfinancial:never）。
- limits — 数字/时间限制（例如maxmeetingminutes、maxcommitusd、quiethours）。
- escalation — 何时上报给人类（家庭、法律、金钱、首次联系等）。
- version + signedat — 用于撤销和更新。
2. 关系合同（成对）
在介绍和清单交换后，存储一份双边协议：
- - 参与方（两个主体/代理）。
- 目的（例如“为项目X安排日程”）。
- 相互约束（时间、渠道、最长会议时长、取消规则）。
- 撤销 — 任何一方均可终止。
保持为双方镜像的小文件或账本条目。
3. 承诺层级
当一个代理与另一个代理对话时，对发言进行分类：
| 层级 | 含义 | 典型要求 |
|---|
| T1 信息性 | 事实、可用性提示 | 清单存在 |
| T2 暂定性 |
“我可以尝试保留…” | 范围允许 |
| T3 确定性 | “我们同意下午3点” | 在限制内 + 关系活跃 |
| T4 约束性 | 金钱/法律/供应商 | 对此操作需明确人类批准 |
如果清单禁止，切勿使用T3/T4语言。
4. 信任/声誉（轻量级）
可选签署的履行与违反承诺日志（不一定是区块链）。用于保守地开始新关系，并随着历史积累放宽上限。
协议互操作性
当技术栈不同时（OpenClaw与其他框架）：
- 1. 公告支持的配置文件（例如“OpenClaw ACP + JSON清单交换”）。
- 协商共享的外部格式（JSON消息 + 清单附件）。
- 优先使用薄通用语言，而非强制某一技术栈的内部细节。
OpenClaw原生映射（概念性）：
- - 身份/网关认证 → 此代理的根信任。
- sessionssend、sessionsspawn、会话日志 → 协调和审计原语。
- 新产品工作可能添加：清单类型、关系记录、承诺事件。
个人代理的语义“能力”
避免模糊标签（“法律帮助”）。优先使用主体批准的任务导向描述：
- - 管辖范围、交付物类型、SLA、数量、上报路径。
匹配委托范围 + 关系目的，而非原始模型基准。
上报层级（始终明确）
- 1. 人类 — 最终权威；敏感或超出范围。
- 代理 — 常规、在清单内、已知关系。
- 拒绝/澄清 — 能力不明确、信任缺失或范围不匹配。
何时加载本技能
- - 用户询问关于代理到代理、两个个人助理、委托、代理间信任，或为此领域在ClawHub上列出技能（概念性回答；除非用户请求终端命令且策略允许，否则不要运行clawhub）。
- 用户希望设计清单、握手或承诺层级——而不仅仅是市场列表。
发布与注册（仅限人类）
要在ClawHub上发布或更新本技能，请使用您机器上的ClawHub CLI（安装、clawhub login，然后在内容变更时使用新的语义版本号运行clawhub publish）。请参阅clawhub.com上的官方ClawHub文档获取当前publish语法和标志——不要依赖此文件内复制粘贴的shell片段。
参考
