When the support agents talk directly: the 2026 inter-agent protocol stack

BLUF: The plumbing for one firm’s AI agent to talk to another firm’s AI agent — no email thread, no phone tree — largely shipped in 2025. Anthropic’s MCP (agent↔tool, Nov 2024), Google’s A2A (agent↔agent, Apr 2025), IBM’s ACP, Cisco’s AGNTCY, and a payments layer (AP2 + x402, Sep 2025) all landed inside 18 months, and by Dec 2025 both MCP and A2A sat under the Linux Foundation with the hyperscalers co-governing. The open question is not whether agents can interoperate — it’s what they say to each other once natural language stops being the constraint, who is liable when a machine-speed negotiation goes wrong, and what an agent-native supply-chain twin must expose to be a first-class participant rather than a scraped data source. This brief maps the stack, the hard problems, and the open questions — it does not draw a TBD-strategy verdict.

Scope note — this is a no-anchor departure brief. Unlike standard RDI outputs, there is no single anchor interview. Like the agentic-prototyping tooling brief, it’s a landscape map assembled from ~20 external sources plus two internal threads that framed the question: [[2026-03-ann-miura-ko|Ann Miura-Ko]] (“B2B purchasing may shift to agent-to-agent transactions”) and [[2026-05-22-intro-dustin-jeremy-ross-jeremy-jawish|Jeremy Jawish/Shift]] (insurers unbundling into purpose-built agents; the liability void). Adjacent internal sources — [[sap-salesforce-enterprise-stickiness-2026-07-08|SAP/Salesforce stickiness]], [[sap-in-semiconductor-supply-chain-2026-05-27|SAP in the semi supply chain]], [[charter]], [[manus-ai-robotics|Manus primer]] — are cited inline where external evidence confirms or complicates them.

Outline changes: none. Executed as the approved 7-section plan. One disambiguation added throughout: “ACP” refers to two different 2025 protocols (IBM’s Agent Communication Protocol and OpenAI/Stripe’s Agentic Commerce Protocol) — flagged wherever it appears.

Per RDI, synthesis is a human activity. This surfaces evidence and ends in questions. Every non-trivial claim carries a source label.

1. The customer-support-to-customer-support channel

Picture the concrete case: a buyer’s support agent and a supplier’s support agent resolving an RMA with no human in the email thread. Today that exchange is “death by a thousand emails” — the exact phrase Greg at NVIDIA used for status data scattered across “SFDC, SAP, the NVIDIA data lake, customer reply-all chains, and 3PL spreadsheets, with no single source of truth” [Interview: Greg + Lonny/NVIDIA, 2026-06-17, via [[sap-salesforce-enterprise-stickiness-2026-07-08]]]. The founder’s framing takes that thread and removes the humans: the buyer-side agent files the return, the supplier-side agent checks warranty status, proposes repair-vs-replace, and both settle a credit — machine to machine.

What actually passes on that channel, in 2026 terms:

A discovery step: the buyer agent finds the supplier’s agent and reads its capabilities from an Agent Card — a JSON document at /.well-known/agent-card.json declaring identity, endpoint, skills, supported modalities, and auth requirements [Public: A2A Protocol Specification, a2a-protocol.org, 2025].
A task object with a defined lifecycle — submitted → working → input-required → auth-required → completed / failed / canceled / rejected — so a long-running RMA stays in sync across both agents rather than living in an inbox [Public: A2A spec, 2025].
Structured artifacts (the return authorization, the warranty determination, the credit memo) rather than prose, though A2A is modality-agnostic and can carry text, files, or forms [Public: Google Developers Blog, A2A announcement, 2025-04-09].
Optionally a payment/settlement leg (the credit or restocking fee), which A2A itself does not handle — that’s the AP2/x402 layer in §6 [Public: Google Cloud Blog, AP2 announcement, 2025-09].

Why TBD cares. The charter already commits to this world: the product is “designed to be invoked by agents” and to present “a structured, legible data model… that both human analysts and LLMs can interact with effectively,” drawing on the Palantir ontology model [Interview/internal: [[charter]]]. If firm-to-firm coordination moves onto agent channels, a supply-chain twin that is only human-readable is invisible to the participants that matter. Ann Miura-Ko named the destination early: “B2B purchasing may shift to agent-to-agent transactions,” with “mathematically-trained purchasers” producing “more efficient pricing” than human psychology allows [Interview: Ann Miura-Ko, 2026-03-06]. External evidence now puts numbers on that intuition — McKinsey estimates agentic commerce could influence $3–5 trillion in global commerce by 2030 [Public: McKinsey via digital.nemko.com, 2025]. The direction Ann sketched in March 2026 is no longer [Speculation]; the infrastructure exists. What remains uncertain is adoption depth and the failure modes (§6).

2. The 2026 protocol stack — who owns which layer

The stack sorted itself into layers faster than it consolidated into one winner. The clean mental model: MCP connects an agent down to tools and data; A2A connects an agent sideways to other agents; a payments layer (AP2/x402/ACP-commerce) lets those agents transact. The competitors to A2A (IBM ACP, ANP, AGNTCY) mostly converged or specialized rather than fought to the death — the most important structural fact of the year.

Protocol	Layer / job	Owner → governance	Transport / mechanism	Adoption signal
MCP (Model Context Protocol)	Agent ↔ tools/data	Anthropic (Nov 2024) → donated to Linux Foundation AAIF, Dec 2025	JSON-RPC; client/server, tools + resources	OpenAI adopted across Agents SDK, Responses API, ChatGPT desktop (Mar 2025); ~97M monthly SDK downloads by Dec 2025; ~28% of Fortune 500 running MCP servers `[Public: Wikipedia/MCP; guptadeepak.com, 2025–2026]`
A2A (Agent2Agent)	Agent ↔ agent	Google (Apr 9, 2025) → donated to Linux Foundation, Jun 23, 2025	JSON-RPC 2.0 over HTTPS + SSE; Agent Cards, task lifecycle	50+ launch partners → 150+ orgs post-donation; founding contributors AWS, Cisco, Microsoft, Salesforce, SAP, ServiceNow `[Public: Linux Foundation press, 2025-06; Google OSS Blog, 2026-04]`
ACP (IBM Agent Communication Protocol)	Agent ↔ agent (REST-native, local-first)	IBM Research / BeeAI (Mar 2025) → Linux Foundation	REST, HTTP/JSON; agent discovery + multimodal messages	Merged with A2A under LF, Aug 29, 2025 — BeeAI agents become A2A-compliant via adapter `[Public: LFAI & Data blog, 2025-08-29]`
ANP (Agent Network Protocol)	Agent ↔ agent, decentralized/open web	Open community; W3C CG white paper	W3C DIDs (`did:wba`), `.well-known/agent-descriptions`, E2E encryption	Positions as “HTTP of the agentic web”; earlier-stage, research/standards track `[Public: agentnetworkprotocol.com; arXiv 2508.00007]`
AGNTCY (“Internet of Agents”)	Full stack: discovery, identity, messaging, observability	Cisco Outshift (Mar 2025) → Linux Foundation, Jul 2025	OASF schema (“DNS for agents”), decentralized Agent Directory, SLIM messaging (quantum-safe)	65+ companies; formative members Cisco, Dell, Google Cloud, Oracle, Red Hat; built with LangChain, Galileo `[Public: Linux Foundation press, 2025-07; Network World, 2025]`

Where they overlap vs. compete:

MCP and A2A are complementary, not rivals — this is now the consensus framing, and both SAP and Salesforce back both [Public: SAP Architecture Center; Salesforce Developers Blog, 2025, via [[sap-salesforce-enterprise-stickiness-2026-07-08]]]. MCP gives an agent hands (tools); A2A gives it peers.
A2A, ACP, and AGNTCY genuinely overlapped on the agent-to-agent layer — and the market resolved it by consolidation under one foundation. IBM’s ACP formally merged into A2A (Aug 2025); AGNTCY joined the Linux Foundation alongside A2A (Jul 2025), positioning as the broader “identity + discovery + observability” envelope around A2A messaging [Synthesis from LFAI blog + LF press, 2025].
ANP is the outlier — decentralized, DID-based, aimed at the open internet rather than the enterprise. If A2A is the enterprise LAN for agents, ANP wants to be the public web. Adoption is far thinner [Synthesis; arXiv 2505.02279, "A survey of agent interoperability protocols"].

The one-year consolidation is the surprise. By December 2025 the two foundational protocols (MCP, A2A) lived under one neutral body (Linux Foundation), governed jointly by companies that compete on everything else — Anthropic donated MCP to the Agentic AI Foundation “backed by AWS, Google, Microsoft, OpenAI, Bloomberg, and Cloudflare” [Public: guptadeepak.com; MCP Wikipedia, Dec 2025]. Standards wars usually take a decade; this one compressed into ~14 months [Synthesis].

3. Beyond human language — what “post-email” agent comms looks like

The founder’s “think beyond human communication” instinct has a 30-year academic lineage and a live 2026 tension: is natural language the interoperability layer, or the fallback? Two camps exist, and 2026 has not resolved which wins.

The structured/semantic lineage (FIPA-ACL). Long before LLMs, the Foundation for Intelligent Physical Agents standardized FIPA-ACL — an agent communication language with ~20+ “performatives” (speech acts like request, inform, propose, agree, refuse), where the performative field encodes intent and the payload carries content [Public: Wikipedia/Agent Communications Language; smythos.com, 2025]. This is the “beyond email” ideal: agents exchange typed, machine-parseable intent, not prose to be re-interpreted. Modern A2A’s task lifecycle and structured artifacts are FIPA-ACL’s descendants [Synthesis].

The natural-language camp. The competing bet is that the LLM is the universal parser, so natural language becomes the interop layer — no schema negotiation needed. Ecma International backed this in December 2025 with the Natural Language Interaction Protocol (NLIP) — five standards (ECMA-430 through 434) plus a technical report, defining “a universal envelope protocol” to replace “hard-coded APIs” for agent-to-agent and human-to-agent exchange [Public: Ecma International, 2025-12-10]. Newer approaches explicitly blend the two — extending FIPA’s performative library “by weaving in natural language alongside JSON” [Public: medium.com/@drajput, 2025].

The machine-speed prize is what “beyond human communication” actually buys. Negotiation is the sharpest example. An English auction that takes 30 minutes with humans “collapses into a few seconds with AI agents”; offers can “update every second based on inventory, SLAs, budgets, and delivery constraints” [Public: keelvar.com; GEP Blog, 2025]. The academic Automated Negotiating Agents Competition (ANAC) has run multi-deal, sequential-opponent negotiation challenges for years and continued in 2025 [Public: arXiv 2604.13914, ANAC 2025]. The B2B breakthrough “isn’t chat-based; it is protocol-based… high-speed exchanges of rich data using multi-round protocols” — optimizing price, payment terms, delivery, volume, service levels, and risk allocation simultaneously [Public: keelvar.com, 2025]. This is precisely Ann Miura-Ko’s “mathematically-trained purchasers” producing “more efficient pricing” — now demonstrable, not hypothetical [Interview: Ann Miura-Ko, 2026-03-06; Public: keelvar/GEP, 2025].

Discovery & capability advertisement — the registry problem. For agents to negotiate, they must first find each other and know what each can do. Four competing discovery mechanisms exist in 2026:

A2A Agent Cards at /.well-known/agent-card.json — the de facto enterprise standard [Public: A2A spec, 2025].
AGNTCY OASF + Agent Directory — a decentralized registry, pitched as “DNS for agents” [Public: Linux Foundation, 2025-07].
ANP .well-known/agent-descriptions with DID-based identity [Public: agentnetworkprotocol.com, 2025].
Vendor galleries — Google Agentspace’s Agent Gallery and Salesforce AgentExchange (§5) — proprietary catalogs inside a platform’s walls [Public: Google Cloud Next 2025; Salesforce, 2025-03-04].

Caution flag — capability advertisement is also an attack surface. Trustwave SpiderLabs showed A2A “could be gamed to route all requests to a rogue AI agent by lying about its capabilities” in its Agent Card [Public: thehackernews.com, 2025-04]. Discovery and deception are the same primitive viewed from two sides (see §6) [Synthesis].

4. Open-source frameworks & standards bodies

Two distinct open-source stacks matter: the orchestration frameworks that build an individual agent, and the protocols/governance bodies that let agents interoperate. Don’t conflate them — LangGraph builds the agent; A2A lets it talk to a CrewAI agent across a firm boundary.

Orchestration frameworks (how you build one agent or a local crew):

Framework	Owner	Orchestration model	Notable trait
LangGraph	LangChain	Directed graphs, conditional edges	Built-in checkpointing / “time travel”; model-agnostic; picked for control + compliance `[Public: langchain.com; QubitTool, 2026]`
AutoGen → AG2	Microsoft → community (AG2)	Conversational GroupChat	Best for agents that negotiate/critique/iterate; model-agnostic `[Public: dev.to/pockit, 2026]`
CrewAI	CrewAI	Role-based “crews,” process types	Lowest learning curve (~20 lines); added native A2A support `[Public: langflow.org; QubitTool, 2026]`
OpenAI Agents SDK	OpenAI	Explicit handoffs	Clean, opinionated; OpenAI-model-bound; adopted MCP `[Public: techsy.io, 2025]`
Claude Agent SDK	Anthropic	Tool-use loop	Claude-model-bound; MCP-native `[Public: QubitTool, 2026]`

These are largely model-agnostic and interoperable via protocols — a CrewAI agent and a LangGraph agent don’t share a runtime; they share A2A [Synthesis]. This maps to the Manus framing of “atomic capabilities” and a “general agent vs. specific task agents” split — the frameworks above are the machinery for building the specific task agents [Interview/primer: [[manus-ai-robotics]], 2025-10-21].

Protocols & governance bodies (how agents interoperate across boundaries):

Google A2A → Linux Foundation (donated Jun 23, 2025; 150+ orgs) [Public: LF press, 2025-06].
Anthropic MCP → Linux Foundation Agentic AI Foundation (Dec 2025) [Public: MCP Wikipedia, 2025].
IBM ACP / BeeAI → Linux Foundation (merged into A2A, Aug 2025) [Public: LFAI blog, 2025-08].
Cisco AGNTCY → Linux Foundation (Jul 2025; 65+ companies) [Public: LF press, 2025-07].
Ecma International NLIP — the natural-language-envelope alternative, a traditional standards body rather than a foundation (Dec 2025) [Public: Ecma, 2025-12].

The pattern: the Linux Foundation became the Switzerland of agent protocols in 2025. Nearly every serious agent-to-agent standard ended up under LF governance within months of launch — a deliberate move to avoid a repeat of the fragmented API-standards wars [Synthesis].

5. Proprietary / platform solutions

Every major enterprise platform shipped an agent layer in 2025, and each is racing to own the identity and registry choke point while professing openness on the wire protocol. They back A2A/MCP publicly, then build a proprietary gallery, identity system, and pricing meter around it. This is the SAP/Salesforce “wrap, don’t replace” instinct applied to agents [Synthesis; cf. [[sap-salesforce-enterprise-stickiness-2026-07-08]]].

Platform	Agent product	Protocol support	Identity model	Pricing model
Salesforce	Agentforce 360 (GA Oct 13, 2025); AgentExchange marketplace (Mar 4, 2025, 200+ partners)	Native MCP client (pilot Jul 2025) + A2A + ACP-commerce (w/ Stripe, Oct 2025)	Salesforce org identity	$2/conversation; Flex Credits ($0.10/action); or $125/user/mo `[Public: Salesforce press, 2025; via [[sap-salesforce-enterprise-stickiness-2026-07-08]]]`
SAP	Joule + Joule Agents (GA Q4 2024); Joule Studio (Q1 2026)	MCP Gateway + A2A Agent Gateway in Integration Suite	SAP BTP identity	AI Units (~$0.40/unit list) `[Public: SAP News; saplicensingexperts.com, 2025]`
Microsoft	Copilot Studio + Azure AI Foundry agents	MCP + A2A	Entra Agent ID — unified agent directory; auto-created for Copilot Studio agents since Mar 18, 2026; Conditional Access + audit	Copilot licensing + consumption `[Public: Microsoft Learn / techcommunity, 2025–2026]`
Google	Agentspace + Agent Gallery; Vertex AI	A2A originator; MCP	Google Cloud IAM	Vertex/Agentspace consumption `[Public: Google Cloud Next 2025]`
NVIDIA	Agent Toolkit / open enterprise agent platform (GTC 2026)	Framework-agnostic; Nemotron models	OpenShell policy/security runtime	Infra/model consumption; 17 named adopters incl. SAP, Salesforce, ServiceNow, Siemens, Cadence, Synopsys, Palantir, Cisco `[Public: VentureBeat; NVIDIA newsroom, 2026, via [[sap-salesforce-enterprise-stickiness-2026-07-08]] & [[sap-in-semiconductor-supply-chain-2026-05-27]]]`

Two observations, not conclusions:

Identity is where the platforms are planting flags. Microsoft’s Entra Agent ID exists to solve “agent sprawl” — agents “created across Copilot Studio, Azure AI Foundry, Power Platform… without IT visibility” [Public: Microsoft techcommunity, 2025]. Whoever owns agent identity owns the audit trail, which (per the liability thread in §6) may be the most defensible layer [Synthesis].
The NVIDIA Agent Toolkit is the supply-chain-relevant one. Its 17 adopters include the EDA and fab-adjacent names (Cadence, Synopsys, Siemens) already in TBD’s orbit — a signal that semiconductor-industry agents will exist inside NVIDIA’s runtime, not only Salesforce’s or SAP’s [Public: VentureBeat, 2026; Synthesis].

6. The hard problems — where it breaks

Every layer that shipped in 2025 has a matching unsolved failure mode in 2026. The infrastructure is real; the safety, identity, and accountability scaffolding is not.

Identity & auth. Agents need verifiable identity across org boundaries, and the field has three incompatible answers: platform-issued (Microsoft Entra Agent ID), decentralized DIDs (ANP’s did:wba, AGNTCY’s crypto identity), and A2A’s per-Agent-Card auth (API key / OAuth / OpenID Connect) [Public: Microsoft Learn; ANP white paper; A2A spec, 2025]. No cross-scheme trust anchor exists yet [Synthesis].

Agent payments & commerce — a crowded, fragile layer. Multiple 2025 protocols compete to let agents transact:

AP2 (Google Cloud + Coinbase, Sep 2025) — gives agents “a wallet, a programmable settlement rail, and auditable proofs,” with the x402 extension (Coinbase, Ethereum Foundation, MetaMask) adding stablecoin settlement [Public: Google Cloud Blog; Coinbase, 2025-09].
ACP-commerce (OpenAI + Stripe, Sep 29, 2025, Apache 2.0) — agent-initiated checkout; powered ChatGPT Instant Checkout. Caution: Instant Checkout was retired in March 2026 after “only about a dozen Shopify merchants ever shipped against it” [Public: Stripe newsroom; search result, 2025–2026]. A live reminder that shipped protocol ≠ adopted protocol [Synthesis].
Naming hazard: “ACP” = IBM’s Agent Communication Protocol AND OpenAI/Stripe’s Agentic Commerce Protocol. Different layers, same acronym [Synthesis].

Liability & accountability — the Jawish void, now industry-wide. Jeremy Jawish gave the sharpest internal articulation: an AI agent “autonomously negotiating injury claims with third-party attorneys” — if it makes an illegal or coercive move, “Is it the insurance company’s fault? Is it Shift’s fault?… Imagine you get a 3 million penalty and it’s because the agent screwed up.” The result: full automation stalls, and “manual human checkpoints remain in place,” a structural bottleneck [Interview: Jeremy Jawish, 2026-05-22]. External evidence confirms this generalizes: automated negotiation research warns that “without policy checks (pricing floors, compliance, data use), agents can ‘optimize’ into unacceptable outcomes” [Public: keelvar.com/GEP, 2025]. Jawish’s insurance-specific void is the same void the whole agent-commerce layer is staring at — nobody has resolved who is liable when two firms’ agents strike a bad machine-speed deal [Synthesis]. This is also where his other thread bites: Shift is “decomposing its five products into discrete agents” (7, targeting 10), and insurers are “actively pivoting away from buying big solutions toward buying individual agents” [Interview: Jeremy Jawish, 2026-05-22] — an unbundling that multiplies the number of agents needing identity, liability coverage, and interop exactly as the accountability model lags.

Semantic mismatch. Even with a shared wire protocol, two agents can disagree on what a field means — the same master-data conflict that plagues SAP↔Salesforce integration today (“which system owns the customer/asset record”), which “middleware moves around but doesn’t resolve” [Synthesis; via [[sap-salesforce-enterprise-stickiness-2026-07-08]]]. A2A standardizes the envelope, not the ontology inside it. This is precisely the gap a shared, legible ontology (the TBD thesis) would fill — flagged as relevance, not conclusion [Synthesis].

Cross-agent prompt injection & security — the leading 2026 risk. MCP has “prompt injection security problems” structurally: tool descriptions “visible to the AI model but not displayed in user interfaces” can carry hidden adversarial instructions [Public: Simon Willison, 2025-04-09]. Invariant Labs demonstrated “cross-server tool shadowing” — one malicious MCP server weaponizing adjacent trusted ones [Public: thehackernews.com, 2025-04]. The Cloud Security Alliance called it an “MCP Security Crisis” of “systemic design flaws” (May 2026) [Public: CSA Labs, 2026-05]. On the A2A side, a compromised agent could “infect others through A2A communication” [Public: arXiv 2506.23260]. Prompt injection is named the leading AI agent security risk of 2026 [Public: techstoriess.com; MDPI Information 17(1):54, 2026]. In the customer-support-to-customer-support scenario, a malicious supplier agent’s Agent Card or tool description is a direct injection vector into the buyer’s agent [Synthesis].

7. Supply-chain-specific note + trends & open questions for TBD

Supply-chain precedent already exists — inter-firm machine-to-machine data exchange is live in automotive, and it isn’t an LLM story. Catena-X runs cross-company data exchange across the auto supply chain using Eclipse Dataspace Connectors (EDC) — standardized connectors that let a supplier’s MES interface with an OEM’s quality platform “while maintaining data sovereignty,” no bilateral integration [Public: internationaldataspaces.org; msg.group, 2025]. Committed participants include Volkswagen, Mercedes-Benz, BMW, Ford, BASF; since April 2025 “Catena-X registration has become integral to BMW’s procurement process” [Public: automotivemanufacturingsolutions.com, 2025]. There is even a Semiconductor-X / Catena-X thread and an OPC-UA integration architecture [Public: catena-x.net whitepaper, 2025-08; via [[sap-in-semiconductor-supply-chain-2026-05-27]]]. The nuance: data spaces solve data sovereignty and exchange (who may see what, under what contract), which is a different and more mature primitive than agent negotiation. A2A-style agents and Catena-X-style data spaces are complementary layers, not competitors — and a supply-chain twin plausibly needs to speak both [Synthesis].

Where 2026 is heading (patterns, not predictions):

Wire-protocol interop is converging (MCP + A2A under LF); the contested frontier is moving up to identity, discovery/registries, payments, and liability [Synthesis].
The platforms are racing to own agent identity and galleries as the new lock-in, even while backing open wire protocols — the same wrap-don’t-replace move seen with data [Synthesis; cf. [[sap-salesforce-enterprise-stickiness-2026-07-08]]].
Adoption is uneven: MCP is genuinely pervasive (~97M downloads); agent-commerce is fragile (Instant Checkout retired after ~12 merchants). “Shipped” and “adopted” are diverging [Public/Synthesis].

Open questions for the founders (end in questions, per RDI):

If firm-to-firm supply-chain coordination moves onto A2A, what exactly should a TBD digital twin expose — an Agent Card + A2A endpoint, an MCP server, a Catena-X EDC connector, or all three? The charter says “invoked by agents”; this is the concrete decision that phrase now implies. Who could tighten it: an NVIDIA IT/SAP architect (Greg’s org); a Catena-X/Cofinity-X integration lead.
Does the semantic-mismatch gap — A2A standardizes the envelope, not the ontology — describe the exact slot a shared legible ontology would occupy, or does each platform’s proprietary data layer (Data 360, SAP Knowledge Graph) close it first? This is the load-bearing question for whether an independent ontology has room to exist. Who: Susan Athey (complements/Trojan-horse framing); a Palantir Foundry architect.
Who becomes liable when two firms’ supply-chain agents strike a bad machine-speed deal — and is that liability void itself a product (parametric coverage for agent-negotiated transactions)? Jawish’s insurance void [Interview: Jeremy Jawish, 2026-05-22] and the parametric-insurance wedge may intersect here. Who: Jawish (follow-up); a specialty/parametric insurer; a reinsurer’s emerging-risk desk.
Is Ann Miura-Ko’s “mathematically-trained purchasers → more efficient pricing” thesis a tailwind or a threat to a chip-pricing/hedging product? If agents negotiate memory contracts at machine speed across many buyers, does that create demand for a neutral price index and hedge, or compete away the inefficiency the wedge relies on? Who: Ann Miura-Ko (follow-up); a DRAM contract buyer.
Which identity anchor wins — platform (Entra Agent ID), decentralized (DID/ANP), or protocol-native (A2A Agent Card auth) — and does TBD need to be issuable/verifiable under whichever the supply chain adopts? Who: a Microsoft Entra or Cisco AGNTCY product contact.
Given Instant Checkout’s retirement after ~12 merchants, what’s the real adoption curve for agent-to-agent commerce in B2B supply chains vs. B2C retail — and which comes first? Who: Stripe/agentic-commerce contact; a B2B procurement platform (GEP, Keelvar).

Confidence summary

Claim area	Confidence	Caveat
Protocol existence, owners, donation dates (MCP, A2A, ACP, AGNTCY, AP2, NLIP)	High	Well-documented, dated, primary-source-backed
MCP/A2A convergence under Linux Foundation	High	Multiple independent sources
Adoption depth (Fortune-500 %, 150+ orgs, 97M downloads)	Medium	Several figures are vendor/advocacy-sourced; directional
Agent-commerce fragility (Instant Checkout retirement)	Medium	From search summary; verify merchant count before external use
Liability void generalizes beyond insurance	Medium–High	Strong internal (Jawish) + external (negotiation-governance) convergence
Semantic-mismatch = ontology opportunity	Low–Medium	`[Synthesis]`/`[Speculation]`; a hypothesis to test, not a finding
Which identity anchor wins	Low	Genuinely unresolved in 2026

Sources

Internal (memory vault)

[[2026-03-ann-miura-ko|Ann Miura-Ko, 2026-03-06]] — agent-to-agent B2B purchasing; mathematical purchasers vs. human psychology
[[2026-05-22-intro-dustin-jeremy-ross-jeremy-jawish|Jeremy Jawish / Shift, 2026-05-22]] — agent unbundling; liability/accountability void
[[sap-salesforce-enterprise-stickiness-2026-07-08|SAP/Salesforce stickiness, 2026-07-08]] — A2A/MCP/AgentExchange/NVIDIA Toolkit; “death by a thousand emails”; wrap-don’t-replace
[[sap-in-semiconductor-supply-chain-2026-05-27|SAP in the semi supply chain, 2026-05-27]] — NVIDIA Agent Toolkit adopters; Catena-X / Semiconductor-X
[[charter]] — “designed to be invoked by agents”; Palantir ontology model
[[manus-ai-robotics|Manus primer, 2025-10-21]] — general vs. task agents; atomic capabilities
[[ai-agentic-prototyping-tooling-2026-06-25]] — format precedent for a no-anchor departure brief

External — protocols & governance

External — discovery, negotiation, FIPA lineage

External — orchestration frameworks

LangChain — best AI agent frameworks 2026 · QubitTool — 2026 framework showdown · dev.to — LangGraph vs CrewAI vs AutoGen (2026)

External — platforms

External — payments & commerce

External — security

Simon Willison — MCP prompt injection (2025-04-09) · The Hacker News — MCP & A2A injection, attack and defense (2025-04) · CSA — MCP Security Crisis (2026-05) · MDPI Information — prompt injection review (2026) · arXiv 2506.23260 — protocol exploits in LLM agent workflows

External — supply-chain data spaces (Catena-X)

International Data Spaces — Catena-X relies on IDS · msg.group — Eclipse Dataspace Connector · AMS — why Catena-X is critical to automotive · Catena-X × OPC-UA architecture whitepaper (2025-08)

Prepared 2026-07-28 via the /research workflow. Landscape map, not a verdict — synthesis is a human activity per RDI.

Project TBD Memory Vault

Explorer

inter-agent-communication-protocols-2026-07-28