Skip to content

Repository files navigation

agent-protocol-standards

A suite of Claude Code skills that encode the open standards for agent identity, delegation, agent-to-agent communication, and agent-to-tool authorization — so that AI personas can verify, formulate, implement, and security-test agentic architectures against what the specs actually say, instead of plausible-but-wrong recall.

Status: design phase complete (2026-08-05); building skill 1 of 8 (agent-credentials-grants) next.

Why this exists

The standards in scope (RFC 8693 token exchange, MCP authorization, A2A, SSF/CAEP, the IETF agent-delegation drafts) are new, fast-moving, and exactly the material where an LLM fabricates specifics — claim names, event-type URIs, delegation nesting order. Each skill replaces fabrication with retrieval: primary-source field tables, copyable artifacts, and a testable validation checklist, with pinned spec revisions and drift disclosure.

The suite (three layers)

Layer Skills Role
Knowledge agent-credentials-grants · agent-delegation-drafts · mcp-authorization · a2a-protocol · revocation-signals What the specs say — facts, artifacts, validation checklists
Judgment agentic-architecture-review Turns spec facts into design verdicts with severity
Deployment openclaw-runtime · enterprise-iam-deployment How to realize the patterns on specific stacks

Documentation map

Read in this order:

  1. agentic-identity-skills-handoff.md — the original brief: objectives, design philosophy, per-skill build briefs, spec URL reference. Some sections superseded — see the status note at its top.
  2. docs/build-decisions.md — the locked design decisions (D1–D9) and their rationale.
  3. docs/skill-development-playbook.mdbinding for every skill: principles P1–P11 with violation tests, the seven stage gates, templates, assertion-ID registry, and the definition of done.
  4. docs/claude-skill-authoring.md — Claude Code skill platform mechanics, verified against live docs (frontmatter, loading, evals, distribution).
  5. docs/roadmap.md — phases, checkpoints, the sprint plan (one skill per sprint: build → test → commit → push), and the status tracker.
  6. docs/qa-strategy.mdbinding accuracy QA: how every skill is verified against the standard it codifies (dual verification, coverage walks, fabrication probes, mutation tests, honest QA stamps). Adversarially reviewed via red team + pre-mortem.

Using the skills

Clone the repo and start Claude Code anywhere inside it — skills in .claude/skills/ load automatically. The primary pattern pairs a persona with a skill:

@architect-reviewer use agent-credentials-grants to review the token flow
@golang-pro use mcp-authorization to implement the resource-server token validation

Review findings and generated code cite assertion IDs (e.g. ACG-007). Every skill shows its pinned standard revisions in a banner; if a newer spec revision exists, you'll be told and you choose whether to proceed. Full consumer guide: playbook Part 5.

Contributing a skill

Follow the playbook's stage gates (Part 3) — a skill that skips a gate doesn't ship. Every skill pins exact spec revisions, carries its own evals with fixtures, and passes scripts/validate-suite.sh locally before it lands. Changes are recorded in CHANGELOG.md.

Acknowledgements

The bundled the-fool skill (structured adversarial review) is from jeffallan/claude-skills, MIT-licensed; it was used for the red-team + pre-mortem review of the QA strategy.

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages