Three Agent Specs Promised for October Remain Unshipped. The Vendor Extensions Filling the Vacuum Are Fragmenting the Stack.
PAP v0.1, the Google Gemini Agent general availability, and the A2A Agent Registry formal spec were all promised for October 2026. As of October 10, none have arrived. The delivery gap is the infrastructure story.
The most significant development in agent infrastructure this month is a silence. October 2026 was the target for three foundational specifications: the Personal Agent Protocol v0.1 , the Google Gemini Agent general availability, and the A2A Agent Registry formal spec. As of October 10, none have shipped. This delivery gap defines the current infrastructure cycle, overshadowing any individual product launch.
The Personal Agent Protocol, championed by Sierra and Meta, was intended to provide a unified identity layer for agentic commerce. Announced October 6 at the Sierra Summit with six launch partners — Shopify, Stripe, Walmart, Genesys, Instinct, and Rocket — and followed on October 9 by a working draft called Poppy with 35 additional design partners including OpenAI, Visa, Mastercard, Okta, and Cloudflare, the protocol defines five building blocks: discovery via a standard /.well-known/poppy.json endpoint, sessions, OAuth-based sign-in, cross-channel session persistence, and task execution across websites, APIs, or company agents. Yet the formal v0.1 specification remains unpublished. Sierra promised it “later in October” along with reference implementation and design workshops. Builders must guess at the final shape of the protocol, choosing between waiting or building against a moving target.
The Google Gemini Agent , announced October 8 at the Gemini at Work 2026 keynote by CEO Thomas Kurian, is positioned as “a single, universal agent for work” — answering questions, handling knowledge work, creating images, and writing code from a single prompt box. It promises persistent coworker agents with their own @agents.company.com email addresses, four kinds of memory (session, semantic, procedural, episodic), and multi-agent orchestration that can run for hours or days. It routes across Google’s Gemini family and Anthropic’s Claude models today, with OpenAI and open-weights models listed as “to follow.” But it remains locked in private preview. Wide availability for Google Workspace customers is “coming soon” on select Business and Enterprise plans. Industry specializations for financial services and legal are in preview; government, healthcare, and retail are “coming soon.” For those outside the Google Cloud ecosystem, the promise of a universal agent remains an abstraction.
The A2A Agent Registry illustrates the risks of indefinite delay most clearly. The proposal, started in June 2025 by A2A maintainer kthota-g, has generated 80-plus comments and community convergence on a Core plus Extensions architecture: a three-layer data model (Agent Card, Publication Record, Trust-Authorization Overlay) and three operations (search/list, resolve, snapshot/reference). But convergence on architecture has not produced a ratified spec. At least five independent implementations are shipping with incompatible APIs — Apicurio Registry, two separate a2a-registry community repos, and a Quarkus LangChain4j integration proposal. As Carles Arnal of Apicurio warned in May 2026 : “at least five independent registry implementations are shipping with incompatible APIs, which means the interoperability window is closing — the longer we wait, the harder it gets to align.” In August 2026, the maintainer redirected the community to AI Catalog and ARD specifications as the official discovery path — newer standards that have their own ratification ahead of them.
This pattern mirrors what we documented with the Docker Agent , where OCI standardizes distribution transport while saying nothing about agent-level trust semantics, permission manifests, or sub-agent guarantees. The same dynamic is playing out at the protocol level: distribution is being solved, but the trust and identity layers underneath remain in flux. Builders making adoption decisions today are choosing based on what ships, not on what is promised.
When formal standards fail to materialize, the industry does not pause. The vacuum is occupied by proprietary vendor extensions and incompatible implementations that embed specific assumptions about identity, trust, and discovery. Those assumptions may never align with the formal specifications once they arrive. The cost of tight coupling to a proprietary extension today will be paid in technical debt tomorrow.
For infrastructure decision-makers, the implication is to treat current vendor-specific registries and gateways as temporary bridges, not long-term foundations. Prioritize modular architectures that allow swapping protocol implementations without a complete rewrite of agent logic. Three things to watch: whether the PAP reference implementation becomes a de facto standard or triggers further fragmentation, whether Google Gemini Agent wide availability arrives before year-end or slips into 2027, and whether the Apicurio RFC PR lands before the five registry implementations diverge past the point of alignment.
Related
The Personal Agent Protocol Is Not a Standard War — It’s the Missing Identity Layer
Google Cloud Ships the First Protocol-Native Agent Gateway
Docker Agent Runs on Every OCI Registry. The Registry Runs the Agent Supply Chain.