The Agent Stack Picks Its Three: MCP, A2A, AP2 — and What the Six-Protocol Era Still Doesn't Solve
A special-edition deep dive on the six wire-format specifications competing to define the agent stack — Model Context Protocol, Agent-to-Agent, AG-UI, A2UI, Agent Payments Protocol, and x402. By mid-2026, three of them are pulling ahead as load-bearing infrastructure. The other three are smaller stories, and the most consequential parts of the picture are the gaps that none of the six, individually, solves.
In this episode
Why three protocols — MCP, A2A, and AP2 — are emerging as the load-bearing layers of the agent stack, and what the adoption, governance, and security evidence actually shows.How the AG-UI / A2UI collision resolved into an interoperability alliance with CopilotKit, Google, and Oracle — and why the AP2 / x402 payments collision is still live, with AP2 increasingly looking like the intent layer above x402's settlement rail.The MCP governance gap: a protocol widely described as an "open standard" with no neutral standards body — and a documented case of unilateral client-side spec divergence by a dominant platform.Why the security research on MCP has moved from theoretical risk to working proofs of concept across four independent teams, with no CVE infrastructure yet to catalog the attacks.The IETF's startling counterpoint: current agent protocol work is "problem-space analysis," not solution-space consensus — and the arXiv proposal for two missing protocol layers above the existing stack.The cross-protocol gaps that none of the six solves: agent identity (now being picked up by a new FIDO Alliance working group), observability (OpenTelemetry GenAI semantic conventions), and liability allocation across multi-protocol agent delegation chains.The practical posture for teams building today: treat MCP and A2A as implementation details, invest in observability and indemnity drafting, and read the FIDO and IETF drafts as they land.Sources & References
Primary anchor
Nate's Newsletter — Agent Protocol Stack: MCP and A2A — the seed framing for this deep diveProtocol specifications & primary announcements
Anthropic — Introducing the Model Context ProtocolMCP Official Specification (2025-06-18)Google Developers Blog — Google Cloud Donates A2A to the Linux FoundationLinux Foundation — Launches the Agent2Agent Protocol ProjectA2A Protocol SpecificationA2A GitHub RepositoryPR Newswire — A2A Protocol Surpasses 150 OrganizationsGoogle Developers Blog — Introducing A2UIA2UI Official SiteCopilotKit — AG-UI and A2UI DistinctionOracle — Agent Spec for A2UI / AG-UIFIDO Alliance — Google Donates Agent Payments Protocol to FIDOGoogle Blog — Agent Payments Protocol & FIDO AllianceAP2 Protocol SpecificationGoogle Cloud Blog — Announcing AP2Coinbase Developer Platform — x402 DocumentationCloudflare Blog — x402 Announcementx402 WhitepaperStandards bodies, governance, and the skeptical case
IETF — draft-yao-catalist-problem-space-analysis — Tier 1, the strongest counterpoint to "settled stack" framingFIDO Alliance — Standards for Trusted AI Agent Interactions — the cross-protocol identity working groupSfeir — La jungle des protocoles IA — "jungle of protocols" landscape analysisarXiv 2511.19699 — Internet of Agents: Layered Architecture (Nov 2025) — Layer 8 / Layer 9 proposalOpenAI Developer Community — MCP Server Guidelines Thread — documented client-side spec divergenceSecurity research
Cloud Security Alliance — Secure Use of AP2 (October 6, 2025) — mandate spoofing + agent coercion threat modelSnyk Labs — Prompt Injection MCPPalo Alto Unit 42 — MCP Attack VectorsInvariant Labs — Tool Poisoning AttacksMicrosoft Developer Blog — Protecting Against Indirect Injection in MCPSimon Willison — MCP Prompt Injection (April 9, 2025)Tenable — MCP Prompt Injection: Not Just for EvilarXiv 2511.07426 — MCP Token Overhead Quantified (Nov 2025)EclipseSource — MCP Context Overload (Jan 22, 2026)Adoption, ecosystem, and enterprise readiness
Digital Applied — MCP Adoption Statistics 2026WorkOS — Everything Your Team Needs to Know About MCP in 2026Google Cloud — What is Model Context ProtocolTyk — Is MCP Dead in 2026? Enterprise Readiness ChecklistGalaxy Research — x402: AI Agents & Crypto Payments — independent x402 structural analysisNevermined — Stablecoin Payments AI Agents Statistics — vendor-aggregated x402 volume figures (treat as directional)Crossmint — Agentic Payments Protocols ComparedObservability standards
Arthur AI — Agentic AI Observability Playbook 2026Braintrust — Agent Observability Complete Guide 2026Fiddler AI — MCP Agent ObservabilityMicrosoft Developer Blog — Build 2026: From Observability to ROILegal & liability
University of Chicago Law Review — Risky Agents Without Intentions — Tier 1 peer-reviewed; the human-in-the-loop oversight standardSSRN Working Paper 5864482 — Consumer Protection Gap for Autonomous Agent PaymentsLathrop GPM — Liability Considerations for Agentic AIJones Walker — AI Vendor Liability SqueezeBraun Miller Law — x402 Legal Framework AnalysisHave questions about this episode? Reach out.
View all episodes
