Agentprivacy agent interop

PVM-V4 Skill — Agent Interoperability

Source: Privacy Value Model V4 + SKILL.md Format Spec + agentskills.io Target context: Agent platform builders, skill file authors, cross-platform compatibility architects, community deployers Architecture: agentprivacy.ai · Sync: sync.soulbis.com · Contact: [email protected]

---

What this is

The agentprivacy architecture is designed to produce portable skill files, not platform-locked configurations. A SKILL.md generated through the spellbook process works on any skills-compatible agent runtime: Claude, Copilot, Codex, Goose, OpenClaw. This skill defines how portability works, what the SKILL.md format guarantees, and how the "launchpad" concept enables communities to deploy their own privacy-first agent configurations.

The SKILL.md format

The SKILL_BASE_FORMAT defines a universal agent skill file:

YAML frontmatter. Machine-readable metadata: name, description, license, compatibility, and a metadata block containing origin, template, alignment, skills coverage, grimoire sources, and verification references. Any compliant runtime can parse this frontmatter to determine whether the skill is compatible with its capabilities.

Markdown body. Human-readable content: origin story, the equation, dual-agent separation, privacy layer, proven vs. conjectured table, role skills, grimoire architecture, and operational principles ("How to Use This Skill").

No runtime dependencies. A SKILL.md has no imports, no API calls, no environment requirements. It is a text file. This is deliberate — the moment a skill requires a specific runtime, it loses portability.

License: CC-BY-SA-4.0. All generated skills are open by default. Share-alike ensures derivative skills carry the same openness. Attribution ensures provenance.

Cross-platform compatibility

The skill file must produce consistent agent behaviour across runtimes that have fundamentally different architectures:

Claude (Anthropic). System prompt + context window. The SKILL.md loads as context that shapes Claude's responses. The privacy layer becomes part of Claude's reasoning frame.

Copilot (Microsoft/GitHub). Code-centric context. The SKILL.md influences code generation, review, and documentation. The equation terms map to code architecture decisions.

Codex/GPT (OpenAI). Function-calling oriented. The skill's operational principles translate to function selection heuristics.

Goose (Block). Developer agent framework. SKILL.md as a workspace configuration that governs how Goose approaches tasks.

OpenClaw. Open-source agent runtime. SKILL.md as a first-class skill file loaded into the agent's skill registry.

Consistency guarantee. The SKILL.md does not guarantee identical outputs across platforms — different runtimes have different capabilities. It guarantees consistent principles: the same equation, the same separation constraint, the same privacy layer, the same proven/conjectured distinction. The how varies by platform; the what and why do not.

The launchpad concept

0xagentprivacy is not a platform — it is a launchpad. Communities deploy their own skill files using the template system:

Community picks a persona. A guild of ZKP researchers chooses the Cipher template. A healthcare privacy group chooses the Healer. A journalism collective chooses the Witness.

Community walks the spellbook. Members work through the template's phases — privacy layer, FPS encounters, grimoire interlocks, reconstruction prompts. Each member generates their own proverbs.

Community generates SKILL.md files. The template produces a skill file for each member. The ~30% immutable DNA (equation, separation, privacy layer) is identical across all members. The ~60% dynamic content (domain application, proverbs, specific role skills) is unique to each member.

Community deploys. The generated SKILL.md files are loaded into whatever agent runtime the community uses. The skill files are portable — the community is not locked to any platform.

Network effects compound. Each community's deployment adds to the network term. Guilds that recognise each other's VRCs expand the anonymity sets in their Privacy Pools. The stratum-weighted network topology means active communities at stratum 3 contribute maximum network value.

Skill composition

Skills are designed to compose. A single agent can load multiple skill files:

Layered loading. Privacy layer skills load first (always). Role skills load by pathway. Multiple role skills can coexist — a Cipher agent that also loads the dark_forest skill has both ZKP and MEV protection capabilities.

Conflict resolution. When two skills recommend contradictory actions, the privacy layer takes precedence. If the privacy layer is silent, the more restrictive recommendation wins (Swordsman principles: default to protection).

Skill discovery. The _index.json registry enables runtimes to discover available skills, their dependencies, and their compatibility. A runtime loading the Cipher skill automatically knows it needs crypto_zkp, personhood_sybil, and academic as role skills.

M(u,y) — the adoption term

The equation term M(u,y) = user sophistication × market maturity captures adoption readiness. Agent interoperability directly affects both factors:

User sophistication (u). A well-designed skill file reduces the sophistication required. The spellbook process educates users as it generates skills. Progressive revelation through the armor tiers means users learn as they advance.

Market maturity (y). Cross-platform compatibility accelerates market maturity by preventing fragmentation. A privacy skill that works everywhere matures faster than one locked to a single platform.

The M(u,y) trap. If M(u,y) = 0, the entire equation zeroes. The most mathematically perfect privacy architecture with zero adoption readiness has zero value. This is the practical urgency of agent interoperability — the architecture must be usable or it is worthless.

Open problems

1. Skill file versioning — how do skills update without breaking deployed agents?
2. Cross-runtime testing — validating that a SKILL.md produces consistent behaviour across platforms.
3. Skill dependency management — when crypto_zkp updates, how do downstream skills like Cipher update?
4. Community skill governance — who decides when a community's skill files need revision?
5. Skill verification — proving that a SKILL.md was generated through the legitimate spellbook process, not fabricated.
6. Runtime capability mapping — how to degrade gracefully when a runtime cannot support all features of a skill file.

---

Verify: agentprivacy.ai · sync.soulbis.com · github.com/mitchuski/agentprivacy-docs