State of the art of the Agent Skills standard and adjacent agent standards, and where this repo's skill model stands against them.
Researched 2026-09-18 in session 4f0a68f9-1435-4150-bb9e-84567dc7db02 (harness-exploration-status). The repo's own skill format: the skill model. Transport protocols: MCP, ACP.
| Aspect | Finding |
|---|---|
| The standard | Open, at agentskills.io (Anthropic, Dec 2025): a folder plus SKILL.md with two required YAML fields (name, description) and a prose body. Optional: license, compatibility, metadata; allowed-tools is experimental. Spec repo agentskills/agentskills: ~40 open issues, no published roadmap. |
| Scope | Packaging only: no versioning, dependencies, structured or compiled format, validation or semantics. No skill IR or formal-verification tooling exists anywhere in the ecosystem, so this repo's format-2 + typed <leaf>.cue model + Z3 model tier has no incumbent and no visible competition. |
| Adoption | Real and cross-harness: SKILL.md is native in OpenAI Codex CLI, Gemini CLI, GitHub Copilot, VS Code, Cursor, ~40 products. A harness-neutral skill IR compiling down to SKILL.md rides this distribution instead of fighting it. |
| Skills Over MCP | SEP-2640, AAIF working group: skills served as read-only MCP Resources under skill:// URIs, with a skill://index.json catalog for progressive disclosure. Early prototypes, yet the only live standards work on skill packaging semantics. Watch it: it shapes what an IR must serialize to (aaif.io/blog/skills-over-mcp). |
Status as of 2026-09.
| Standard | Status | For this stack |
|---|---|---|
| AGENTS.md | A deliberately spec-free markdown convention, AAIF-stewarded, 60k+ repos. | Refuses structure; nothing to build on. |
| Evals | No open standard exists or incubates (AAIF has none). The one convergent pattern is the OpenAI Codex recipe: prompt → captured JSONL trace (codex exec --json) → deterministic checks plus schema-constrained rubric grading. | Structurally identical to this repo's skilleval. Aligning vocabulary is cheap insurance; there is no format to conform to. |
| Codex hooks | Shipped, feature-flagged hooks.json: shell scripts at loop points, a direct analogue of Claude Code hooks. Its exec --json JSONL is a de-facto transcript format. | No cross-harness hook or transcript standard, so supervision frontends stay per-harness. |
| A2A v1.0/v1.0.1 | Agent ↔ agent, Linux Foundation, 150+ orgs. | A different layer; irrelevant until multiple cooperating agents are fronted. |
| Agent identity and auth | Pre-standard: two individual IETF drafts (not WG-adopted), OpenID AuthZEN framing; multi-hop delegation unsolved. | Build nothing on it. |
| WebMCP, K8s agent-sandbox, payments, OWASP ANS, Agntcy | WebMCP is a W3C CG draft in a Chrome origin trial; payments are AP2/x402/Agentic Commerce. | Noise. |
Standards cover transport (MCP, ACP) and packaging (SKILL.md). Execution semantics and evaluation have zero incumbent and zero incubation: exactly the layers built here (skill IR, Z3 semantics, skilleval). Build them, keep SKILL.md as the compile target, track SEP-2640.