Standards¶
Open standards and conventions shaping the AI agent ecosystem.
Pages¶
- A2UI: Framework-Agnostic Generative UI Standard — An open standard for agents to emit declarative UI blueprints that a host application renders with its own native components
- ACDL: A Language for Describing Agentic LLM Contexts — A formal notation for specifying how an LLM agent's context is assembled and evolves across interaction steps
- Agent-to-Agent (A2A) Protocol — An open standard for inter-agent communication — agents discover capabilities, delegate tasks, and exchange results over HTTP
- Agent Cards: Capability Discovery Standard — Machine-readable JSON descriptors that advertise agent capabilities, skills, and authentication at a well-known URL
- Agent Definition Formats: How Tools Define Agent Behavior — Agent definitions control system prompt, tool access, model selection, and permissions — the format varies by tool but the concerns are the same across all implementations
- Agent Skills: Cross-Tool Task Knowledge Standard — The Agent Skills open standard packages task-specific knowledge into portable SKILL.md folders that AI coding tools can discover and load on demand
- Agentic Resource Discovery: Federated Pre-Invocation Search — A v0.9 draft open spec defining ai-catalog.json manifests and federated registries so agents find MCP, A2A, and skills before invocation
- AGENTS.md: A README for AI Coding Agents — AGENTS.md is an open standard for a project-level instruction file that gives AI coding agents the context they need to work effectively in a codebase
- Directory-Aware Plugin Suggestions via
pluginSuggestionMarketplaces— A three-lever managed-settings pattern that pins only directory-relevant plugins from admin-allowlisted marketplaces in the Claude Code/pluginDiscover tab - llms.txt: Making Your Project Discoverable to AI Agents — Publish a standardized /llms.txt so agents can navigate your site without crawling every page
- MCP: The Plumbing Behind Agent Tool Access — The Model Context Protocol is an open standard for connecting AI agents to external tools and data sources
- OAuth Client ID Metadata Documents (CIMD) for MCP Servers — CIMD makes an OAuth client_id a URL that dereferences to a JSON metadata document — the registration mechanism MCP recommends for cloud-hosted servers
- OpenAPI as Agent Tool Specification — Use existing OpenAPI specs as the source of truth for agent tool definitions instead of writing schemas by hand
- OpenTelemetry for Agent Observability — Vendor-neutral tracing standard for LLM calls, tool invocations, and agent spans using GenAI semantic conventions
- Plugin and Extension Packaging: Distributing Agent Capabilities — Package agents, skills, MCP servers, and hooks into installable bundles that solve the distribution problem for agent capabilities
- Portable Agent Definitions: Full-Stack Identity as Code — Package an entire agent as a version-controlled, framework-agnostic git artifact using the gitagent open standard
- SUDP: Secret-Use Delegation Protocol for Agentic Systems — A proposed three-role protocol — requester proposes, user authorizes, custodian redeems once — so an agent never holds reusable authority for a user's secrets
- Swarm Skills: Multi-Agent Extension of the Agent Skills Standard — A May 2026 proposal that extends Anthropic's Agent Skills with multi-agent metadata, a per-role file layout, and a self-evolution lifecycle so coordination protocols travel as portable artifacts
- Symphony: Open Spec for Issue-Tracker-Driven Coding Agent Orchestration — Symphony is OpenAI's open spec for orchestrating coding agents from an issue tracker — each open ticket becomes an isolated workspace running its own agent session
- Tool Calling Schema Standards — Tool definitions converge on JSON Schema across providers but differ in field names, strict modes, and wrapping structures
- WebMCP: Browser-Hosted Tool Contracts for In-Page AI Agents — A W3C Community Group draft that lets a web page register JavaScript tools through
navigator.modelContextso an in-tab browser agent calls them by name instead of actuating the DOM