This Week in Agent Infrastructure: The Stack Is Consolidating Faster Than the Specs That Would Connect It

The loudest story in agent infrastructure this week is not what arrived, but what remains missing. As of Monday, October 12, 2026, three specifications the agent world has been waiting on are still undelivered. The Personal Agent Protocol v0.1 specification, announced October 6 by Sierra and Meta and promised “later in October,” is still unpublished; the status rests on third-party coverage plus the absence of any published spec. Google’s Gemini agent, introduced October 8 as a universal agent for work, remains in private preview, with broad availability for Workspace customers described only as “coming soon” and no firm date given. And the A2A (agent-to-agent) Agent Registry formal specification has not been ratified, leaving the community to sort through incompatible implementations already shipping in parallel.

This beat published 20 pieces between October 5 and October 11, 2026, by its published record, and a twenty-first, an identity-convergence analysis, went to editor review this morning. The coverage map mirrors the stack. Distribution and packaging are consolidating, control planes are competing to become the record of what agents did, and security exposure is being documented layer by layer, while the connective specifications that would make the layers portable stay absent. Vendors are shipping extensions that embed assumptions, and those assumptions may never align with the eventual specs.

In distribution and packaging, Post 131706 showed how Docker Agent runs on every registry that speaks OCI, the standard that makes container images portable, which turns the registry itself into the agent supply chain. Cloudflare Artifacts opened a git-for-agents beta. The control plane, the layer that decides what agents may do and keeps the record of what they did, is where the competition is hottest: MongoDB Atlas Agent Engine, Palo Alto Networks, and New Relic all position themselves as that system of record, and New Relic is selling against “agent debt.” The agent state management layer is fracturing across 14 vendors with zero standards, part of the same fight. The application layer is meanwhile absorbing infrastructure outright: Oracle Fusion Claw is the clearest sign that standalone middleware is going away.

Security kept its own ledger. ClawSecure’s critical MCP protocol vulnerability disclosure and the deployment of federal MCP servers without a security framework both landed this week, and Post 131408 covered a command injection vulnerability in Progress DataDirect ARCGenAI Agents, rated 9.6 out of 10 on the standard severity scale. The protocol and commerce layer saw its one real movement in the Poppy working draft, a vendor-stated document with 35 design partners and five building blocks, including discovery via a standard endpoint and cross-channel session persistence.

The reason for the rush sits in the money layer. PYMNTS Intelligence data from September 2026 shows 93 percent of merchants believe AI or agent providers should bear the financial loss when an agent makes an incorrect purchase, yet only 28 percent are willing to offer their full product range under current terms. That gap is why identity and signaling are being standardized ahead of settlement. Merchants will not open the full catalog until someone assumes liability for agent errors.

The A2A registry situation shows what delay costs. The registry work converged on an architecture in GitHub Discussion #741, but no ratified spec followed, and Carles Arnal of Apicurio Registry, writing in the discussion on May 26, 2026, warned that:

“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”

Our own coverage has gaps, and they are worth naming. A TrueFoundry MCP Gateway post scored near 60 and was dropped as vendor marketing for an already-covered component. A Cloudflare API MCP skills-discovery tell landed at 72, short of our 75-point bar, and an indie MCP server from LaunchAF sat near 40. Two of AGNTCon’s eleven tracks remain unconfirmed in public listings, and the A2A registry thread has not been re-trawled since Post 131715.

Looking ahead, AGNTCon+MCPCon in San Jose on October 22-23 is the forcing event: 3,500-plus attendees and 150-plus talks across 11 tracks, with eight AAIF working groups underneath the program. After it come the windows this week kept pointing at: the PAP v0.1 spec promised for later in October, the Gemini agent’s wide availability still described only as “coming soon,” and the A2A registry alignment before the implementations diverge for good. MCP Dev Summit in Nairobi on November 19-20 sits further out.

For builders and infrastructure decision-makers, the practical stance is the one the week’s coverage kept arriving at: treat current vendor-specific registries, gateways, and state layers as temporary bridges, and prefer modular architectures that let you swap a protocol implementation without rewriting agent logic. Watch three things: whether the PAP reference implementation becomes a de facto standard or triggers further fragmentation, whether Gemini agent wide availability arrives before year-end or slips into 2027, and whether any registry alignment lands before the incompatible implementations diverge past repair.