Top AI Agent Standards to Know in 2026

Protocols tell agents how to connect. Standards tell them what to know. As the agent ecosystem matures, a second layer of convergence is emerging: open formats that give agents consistent, structured context — about projects, capabilities, and design systems. Unlike protocols (which define communication between systems), these standards are file-based, human-readable, and version-controlled alongside your code. Here are the three standards shaping how agents are informed and extended in 2026.

1. AGENTS.md

agentsmd/agents.md | Agentic AI Foundation (Linux Foundation) | MIT

The universal context file for AI coding agents. Where a README explains a project to humans, AGENTS.md explains it to agents: build commands, test commands, code style conventions, testing frameworks, architectural decisions, and anything else an agent needs to work effectively in the codebase. Plain Markdown, no required schema, no tooling to install — any agent that reads it benefits immediately.

The problem it solves is fragmentation. Before AGENTS.md, every tool was reading different files, or nothing. Cursor read .cursorrules. Claude read CLAUDE.md. Most agents read whatever they found and hoped for the best. AGENTS.md gives a single, predictable location for agent-specific context without bloating the human README with instructions no human needs.

Adopted by OpenAI Codex, Cursor, GitHub Copilot, and others — reported across 60,000+ open-source repositories as of mid-2026. Governance moved to the Agentic AI Foundation (AAIF) under the Linux Foundation, the same body that now stewards MCP.

Avoiding vendor lock-in: The pragmatic pattern many teams use is to write AGENTS.md as the canonical source of truth, then in tool-specific files (like CLAUDE.md or .cursorrules) simply instruct the agent to read AGENTS.md. One file to maintain, every tool benefits. If a tool stops being used, nothing is lost.

2. Agent Skills (SKILL.md)

agentskills.io | Open standard | MIT

Where AGENTS.md tells agents what a project is, Agent Skills tell agents how to do something — and crucially, that capability travels with the agent across any project. A skill is a folder containing a SKILL.md file with two required YAML fields (name and description) and a Markdown body with instructions. Optional assets like scripts, templates, and reference files live alongside it.

The distinction matters: AGENTS.md is project-scoped context. Agent Skills are reusable, portable capabilities — domain expertise, team-specific workflows, and repeatable procedures that agents load on demand. A skill for writing commit messages, one for generating migration scripts, one for running the company's deployment checklist: each is self-contained, version-controlled, and usable in any compatible agent.

Originally developed at Anthropic and released as an open standard in late 2025, it has since been adopted by Claude Code, OpenAI Codex, Cursor, VS Code, and reported 30+ other tools. Partners including Atlassian, Figma, Stripe, and Notion published skills at launch.

The format is intentionally minimal. Two required fields and a Markdown body — simple enough to implement in an afternoon. No protocol negotiation, no runtime dependencies, no auth flows.

The open format is what makes marketplaces possible. Because a skill is just a folder with a Markdown file, anyone can publish one and any compatible agent can consume it. skills.sh by Vercel is the most active registry today, hosting skills from Anthropic, GitHub, OpenAI, and the community — installable with a single npx skills add <owner/repo> command. The existence of a thriving marketplace is the clearest signal that the standard is working.

3. DESIGN.md

google-labs-code/design.md | Google Labs | Apache-2.0 | ⭐ ~14.6k

The newest of the three, and the most specialized. DESIGN.md is a format for encoding a project's visual identity system in a way agents can read and apply when generating UI code. It combines machine-readable design tokens (YAML) with human-readable rationale (Markdown prose) — giving agents not just the values but the reasoning behind them.

Without something like DESIGN.md, agents generating frontend code have no reliable way to know a brand's colors, typography scale, spacing system, or interaction patterns. They guess from comments in CSS files, or they ignore design consistency entirely. DESIGN.md solves this by making the design system a first-class input to the agent.

Google Labs introduced and open-sourced DESIGN.md as the export/import format for Google Stitch, an AI design canvas that uses Gemini to generate UI from natural language. Designers export a DESIGN.md from Stitch; developers import it into their project; agents use it to keep generated code on-brand. An npm package handles validation (npx @google/design.md lint DESIGN.md), diffing, and export to Tailwind, CSS variables, or W3C Design Token format.

Status: currently in alpha. The format is still evolving and breaking changes are expected. Worth watching and experimenting with, but not yet a safe dependency for production workflows.

Special Mentions

CLI tools as a convention: Not a formal standard, but worth naming. As agents become more capable of invoking shell commands, CLI interfaces for well-known tools (git, gh, docker) are increasingly the simpler, cheaper alternative to a full MCP server — when the tool is already well-documented and the agent can reason about flags. The MCP vs CLI debate has been heated; the practical answer is: MCP for APIs and internal systems, CLI for tools that already have mature interfaces. The AI Agent Roadmap covers this tradeoff in Phase 4.

Key Takeaways

Three complementary layers: AGENTS.md gives agents project context; SKILL.md gives them portable capabilities; DESIGN.md gives them visual identity. They solve different problems and work together.
Markdown as the lingua franca: All three formats are Markdown-first, human-readable, and version-control-friendly. No complex serialization, no runtime dependencies.
Open foundations matter: AGENTS.md is now AAIF-governed alongside MCP, signaling the industry wants neutral, community-owned standards for this layer too.
One file, many tools: The lock-in pattern to avoid is maintaining separate context files per tool. AGENTS.md as the canonical source, referenced from tool-specific files, is the practical answer.

What to Read Next

Enjoying content like this? Sign up for Agent Briefings, where I share insights and news on building and scaling AI agents.