Summary

This page is a watchlist for draft and early-stage standards in the agent ecosystem. Everything here is unstable: pre-consensus specs, origin trials, single-vendor proposals, or specs defined by adoption rather than a pinned version. Track them; do not build production dependencies on them without a concrete reason.

Standards that have stabilized enough to earn dedicated coverage no longer live here - they moved to their canonical homes and are listed in "Graduated to stable coverage" below.

Watch, do not depend without a concrete compatibility reason: , A2UI, AG-UI, OpenUI, AIPREF, , AGNTCY/OASF,
Agentic commerce and payment protocols (, , , , , and card-network schemes) moved to their own section - see Agentic Commerce
Everything on this page carries a status and a date so you can judge staleness

Graduated to Stable Coverage

These were once tracked here as emerging. They are now stable enough to have canonical homes elsewhere in Agent Surface. Follow the pointer for the full treatment; this page no longer re-explains them.

workflows (stable at v1.1.0) → API Surface: Arazzo Workflows
OpenTelemetry GenAI semantic conventions (still Development, but the practical observability baseline) → Testing: Observability
(Linux Foundation context convention) → Context Files: AGENTS.md
/ llms-full.txt → Discovery: llms.txt
and schema.org structured data → Discovery: Structured Data
(still a draft, but substantial enough to need implementation guidance) → Protocols: WebMCP

Draft & Emerging Watchlist

Web Bot Auth - Signed Bot Identity

An IETF effort: the WebBotAuth working group is chartered, building on RFC 9421 HTTP Message Signatures so that an accountable agent can cryptographically sign its requests and a site can distinguish it from anonymous automation. Cloudflare has run it in production since March 2026, which gives it real deployment signal ahead of a finished RFC. Relevant to anyone deciding how to let "good bots" through bot-management controls.

A2UI v0.9 - Generative UI Over Agent Transports

A Google generative-UI specification (v0.9, July 2026) for agents that emit UI rather than only text. It is designed to ride existing transports - AG-UI, , and - rather than define a new one. Extremely new: a sub-1.0 spec with essentially no production track record. Track it as a signal of where generative-UI standardization is heading.

AG-UI - Adoption Without a Pinned Version

AG-UI is an agent-to-UI event protocol that originated at CopilotKit and has no canonical pinned spec version - its status is best read through its adoption. Implementations span Microsoft Agent Framework, Google ADK, AWS Strands, Mastra, Pydantic AI, LangGraph, and CrewAI. The multi-vendor SDK support makes it the current de facto layer for streaming agent state to a UI, but the absence of a versioned specification means the contract can shift under you. Adopt via a maintained SDK, not by hand-implementing the wire format.

OpenUI - Vendor-Authored Generative-UI Standard

OpenUI (openui.com) is an open generative-UI standard from Thesys, MIT-licensed and unversioned: a compact streaming UI language that a model emits as a component tree, plus a renderer-agnostic runtime with first-party React support and community Vue and Svelte ports. Its pitch is efficiency against the obvious alternative - Thesys claims roughly 3x faster rendering and about 67% fewer than streaming equivalent JSON - which is a real problem, since JSON component trees are an expensive way to spend an output budget.

It does not define its own transport. Framework integrations ride AG-UI, the entry above: an AG-UI adapter emits the event stream and OpenUI's client-side agUIAdapter() parses it, the same layering A2UI uses. That makes it a peer of A2UI rather than a competitor to AG-UI, and it means AG-UI's missing version pin is inherited here too.

Status: single-vendor, no published spec version, real download volume but no cross-vendor governance. Watch it as the second serious attempt to standardize what an agent emits when it emits UI; if you adopt it, pin the packages yourself, because upstream ships unpinned.

AIPREF - Content-Usage Preferences

An IETF effort to standardize a Content-Usage HTTP header (and companion signals) letting publishers express how their content may be used for AI. It is pre-consensus: the working group is still shaping the vocabulary, and nothing is settled. Watch it alongside as the AI-usage-preference space converges.

NLWeb - Structured Retrieval

NLWeb provides server-side natural-language retrieval over structured site content. Its reference implementation supports MCP integration. WebMCP exposes page-local browser tools and serves a different execution context. Assess NLWeb against your retrieval needs and intended clients; see catalogs, feeds, and query verification. Evaluate feed , query correctness, and streaming behavior separately.

AGNTCY / OASF - Multi-Agent Infrastructure and Capability Discovery

AGNTCY is a Linux Foundation project (donated by Cisco, July 2025) building shared infrastructure for cross-organization agent systems: discovery, identity, messaging (SLIM), and observability. Cisco, Dell, Google Cloud, Oracle, and Red Hat are formative members, with 65+ supporting companies.

OASF (Open Agent Schema Framework) is AGNTCY's discovery component: normalized schemas for describing agent and tool capabilities, with the goal of making capability discovery work like a package registry - search by what a tool does, not who publishes it. OASF is at v1.2.0 and integrates with Oracle's Open Agent Specification, but the schemas have not converged on a stable version and the surrounding tooling (search indexes, registry operators) remains nascent. A spec to contribute to and monitor.

AGNTCY Dir is the registry/tooling half - the and gateway agents use to publish and query OASF records. Dir v1.7.0 (2026-08-18) added an HTTP POST /v1/search endpoint with parity to the dirctl search CLI, so a caller no longer needs a gRPC client, plus a POST /v1/extract endpoint that pulls OASF out of free-form text using the same extractor the CLI and gateway share. Dir v1.7.1 (2026-09-22) is a stability release: an install manifest becomes the source of truth for package state (so uninstall no longer requires contacting the server), bare-name lookups resolve to the highest version instead of the most recently pushed one, and it adds upgrade/prune/outdated/pin-unpin lifecycle commands. Same maturity read as OASF above: real tooling, not yet something to build a production dependency on.

ANP - W3C AI Agent Protocol

ANP, from the W3C AI Agent Protocol Community Group, targets the problem MCP and A2A leave open: how agents establish trusted identity and communicate when they share no central authority. It combines Decentralized Identifiers (), JSON-LD capability metadata, and end-to-end encryption. It remains a draft with published working drafts but no production reference implementation. Worth monitoring for cross-organization agent networks where no single provider can anchor trust.

Agentic Commerce - Moved

Agent-initiated commerce and payment protocols - UCP, ACP, AP2, x402, MPP, and the Visa and Mastercard card-network agent schemes (Trusted Agent Protocol, Agent Pay, Verifiable Intent) - have their own section: Agentic Commerce. They no longer live on this watchlist; the card-network schemes have real, sourced backing (Visa + Cloudflare, Mastercard + Google) even though they remain single-network efforts without the cross-vendor governance UCP or AP2 have.

Protocol Comparison - when to use MCP, A2A, Arazzo, ANP, and OASF
MCP Protocol - the stable foundation for tool connectivity
A2A Protocol - agent-to-agent delegation
Agentic Commerce - UCP, ACP, payments, and trust and risk
Discovery: Well-Known Endpoints - where discovery documents live today

Emerging Standards Watchlist