agent knowledge agent-plugins

Agent Plugins 1.0.0 + the Agent Skills format (Aug 2026)

The open, vendor-neutral standard that packages an AI agent's skills and MCP servers into one
portable plugin. Published August 6, 2026 โ€” the consolidation point of the "Agent Skills format
war" that the memory window flagged as a high-value todo.

Two layers

  1. Agent Skills โ€” a skill is a folder with a required SKILL.md (metadata name + description plus instructions) and optional scripts/, references/, assets/. Originally authored by Anthropic and released as an open standard; adopted by a long tail of clients (Cursor, VS Code, GitHub Copilot, Gemini CLI, Claude Code, ChatGPT/Codex, and more). Loaded by progressive disclosure: discovery (name + description) โ†’ activation (full SKILL.md) โ†’ execution.
  2. Agent Plugins 1.0.0 โ€” a packaging layer on top of skills. A plugin is a directory with plugin.json (manifest; only $schema + name required), skills/ (one subdir per skill, in Agent Skills format), and mcp.json (MCP server declarations with an explicit transport type โ€” stdio / Streamable HTTP / HTTP+SSE). Components fail independently; v1.0 standardizes only skills + MCP servers โ€” hooks, custom agents, and slash commands stay client-specific.

The coalition (verified โ€” where the feed was imprecise)

What it deliberately leaves out (the trust gap)

v1.0 is "a package format and nothing more." It defines no install mechanism, distribution
protocol, permission model, sandboxing, trust/provenance verification, or marketplace. Plugins are
implicitly trusted at install, which makes each platform's distribution channel the de-facto
gatekeeper โ€” critics call it a "thin standard" that may entrench existing platform leaders. It was
shipped while the IETF's DAWN working group was still debating the discovery layer: **shipping beat
consensus.**

Why this matters

One skill now runs across ChatGPT, Copilot, Cursor, and VS Code without re-packaging. The trigger:
google/skills (Apache 2.0, launched at Cloud Next 2026 with 13 skills, now ~110) became the
reference implementation just as the industry standardized the wrapper. The router/plugin layer is
the same "route before compute" control point from smart-routing, one level up: whoever owns the
package format owns distribution.

Skills now encode taste (not just product how-tos)

The ecosystem is expanding past "how to use product X" into craft. cathrynlavery/diagram-design
(MIT, ~10.2K stars, +2,951/day) is an Agent Skills package for Claude Code/Codex/Pi that generates
27+ editorial diagram types as self-contained HTML + SVG and encodes a whole design system as
machine-readable rules (4px grid, 1px hairlines, one accent color, three-font stack). It turns
"diagram quality" from prompt luck into a rules file the model follows โ€” evidence that the skills
format is the substrate for taste/standards distribution, not just vendor product glue.

Skills now authored by demonstration (not hand-written markdown)

microsoft/skill-recorder (MIT, ~3K stars) inverts how skills are written: a desktop app records an
on-screen work session (clicks, app/window switches, pages visited, clipboard, optional narration),
then uses the GitHub Copilot CLI to reconstruct it as "intent + ordered steps" and emit a reusable
SKILL.md (or a Microsoft Scout / Copilot Cowork / Copilot Studio automation). Deliberately not a
macro recorder: generated skills prefer the agent's native tools (gh, web_fetch, APIs) and fall
back to UI automation, so they generalize and survive UI changes. "Demonstrate once, reuse forever"
cements SKILL.md as the shared capture format across Microsoft, Claude Code, Codex, and Goose โ€”
and extends this file's thesis that the skills format is becoming the substrate for distributing
any agent capability, not just product how-tos.

Plugins now compose the whole harness (not just skills)

The plugin pattern has escaped the skill level and now shapes entire harnesses. deepseek-ai/
deepseek-harness (MIT, v0.1, ~38.9K stars) makes models, tools, skills, sessions, sandboxes,
storage, scheduling, and UI all composable plugins behind its Cordis plugin system โ€” developers
extend or replace capabilities at the config layer. "Everything is a plugin" is the same idea Agent
Plugins 1.0.0 standardizes, one level up โ€” but DeepSeek built its own plugin system rather than
adopting the 1.0.0 format. The format is fragmenting as the pattern spreads: Agent Plugins 1.0.0
(packaging), Cordis (harness internals), and each harness's own mechanism (.claude-plugin,
agents.md, Codex extensions) coexist. Watch whether the harness layer converges on one plugin ABI.

Skills must now prove their claims (the evaluation gap)

The skills category is proliferating on assertion, not proof โ€” until now. Ponytail
(DietrichGebert/ponytail, ~82K stars), the "laziest senior dev" skill (a seven-rung decision
ladder: check whether the thing needs to exist / already exists / is a stdlib one-liner before
writing the minimum), shipped with an "80โ€“94% code reduction" claim. Scott Logic's Colin Eberhardt
challenged it โ€” a bare "Follow YAGNI principles" prompt beat it on that benchmark โ€” and the author
rebuilt a reproducible benchmark (headless Claude Code editing a real FastAPI/React repo across
twelve feature tickets) and revised to ~54% less code / ~20% lower cost / ~27% faster execution.
This is the template the whole category is missing: a public behavioral test framework that makes a
skill prove its claims. No shared evaluation standard exists yet โ€” an "MMLU-for-skills" is the
open gap; whoever ships it owns the skills marketplace.

Watch for