Agentic AI Foundation [What It Changes]

TL;DR

• The Agentic AI Foundation is the Linux Foundation directed fund that now hosts MCP, goose, AGENTS.md, agentgateway, and — as of August 17, 2026 — Google's Agent2Agent protocol under one neutral roof.

• A2A is the shared rulebook for one agent discovering another, handing off a task, and getting a result back; MCP remains the rulebook for an agent talking to a tool, file, or API.

• Cloud vendors already list A2A support in Google Cloud, Azure AI Foundry, and AWS Bedrock AgentCore, so a small shop buying agent software should ask which A2A revision and which bindings a vendor actually ships.

• Neutral governance does not write your access policy, your audit log, or your tenant isolation; those still sit on the operator who lets a partner agent touch a job.

Key Takeaways

• Treat "Agentic AI Foundation" as a governance home, not a product SKU: the specifications stay separate, with their own maintainers and release calendars.

• Ask vendors three buyer questions: which A2A revision, how they treat unsigned agent cards, and who keeps the audit trail when a task crosses a company line.

• A two-truck HVAC shop, a 10-person agency, and a solo clinic feel this first as fewer custom handoffs between the tools they already run, not as a new dashboard they have to staff.

• If you already automate intake, dispatch, or document routing, the practical move is to add a card and a conformance check — not to rip out the workflow.

What the Agentic AI Foundation is

The Agentic AI Foundation is the Linux Foundation directed fund that gives the open agent stack — tools, runtimes, instruction files, gateways, and now agent-to-agent talk — a vendor-neutral home so a job can move between systems without a one-off integration for every pairing.

That sentence is the whole point for a small operator. A two-truck HVAC company does not care which foundation hosts a protocol. It cares that the dispatch board, the parts vendor's bot, and the after-hours answering service can pass a work order without a person retyping the address. A 10-person marketing agency cares that a form capture tool can hand a qualified lead to a fulfillment agent the client already pays for. A solo-run clinic cares that an intake agent can ask a billing agent for a claim status without exporting a spreadsheet at 9 p.m.

Until this year, each of those handoffs was a custom job. The vendor of the first agent and the vendor of the second agent did not share a discovery format, a task object, or a way to prove which bot was speaking. The Agentic AI Foundation is the industry's attempt to put those rules in a place no single cloud owns. The Linux Foundation announced the fund on December 9, 2025, with founding projects from Anthropic, Block, and OpenAI, and with platinum members that include Amazon Web Services, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI.

This hub explains what joined the foundation as of August 17, 2026, how the pieces fit, what a small team should demand from a vendor, and what the transfer still does not settle. Teams comparing this stack against day-to-day automation can also read the state of small-business automation for the operator-level map this protocol now sits under.

What joined on August 17

On August 17, 2026, the Agentic AI Foundation said Agent2Agent (A2A) is joining as a hosted project. A2A is the open standard for how agents discover each other, delegate tasks, and exchange work across frameworks and vendors. According to the Agentic AI Foundation, A2A is backed by over 150 organizations and is already running in production across supply chain, financial services, and mobile platforms. A2A lists 150-plus backing organizations.

Forbes reported the same transfer on August 19, 2026, and a Yahoo syndication of that column carried the same text later that day. The framing in that column matters: Google had already donated A2A to the Linux Foundation in June 2025. What changed in August 2026 is which foundation hosts the work, not who privately controls the spec.

According to Forbes, Google launched A2A in April 2025 and the protocol reached v1.0 in March 2026, adding multi-protocol bindings, version negotiation, multi-tenancy, and cryptographically signed agent cards. A2A v1.0 shipped in March 2026. The canonical specification now publishes latest released version 1.0.0.

The foundation itself is young. OpenAI's co-founding note dated December 9, 2025 describes AAIF as a directed fund under the Linux Foundation, built so agent infrastructure is not trapped in one company's roadmap. The public face of that fund is aaif.io, and the GitHub org at github.com/aaif lists the hosted projects and working groups.

How A2A and MCP actually split the work

Plain language first. MCP is how an agent reaches a tool. A2A is how one agent asks another agent to finish a job.

The A2A project site states that split without hedging: MCP standardizes agent-to-tool communication; A2A standardizes agent-to-agent communication. They are complementary. You equip one agent with MCP so it can read a calendar, a CRM, or a parts catalog. You use A2A so that equipped agent can hand a task to a specialist agent you do not own.

Google's original April 9, 2025 launch post described the mechanism as a client agent and a remote agent. The remote agent publishes an Agent Card — a JSON document that says who it is, what it can do, where to reach it, and how to authenticate. The client reads the card, sends a task, and gets back a result the spec calls an artifact. Tasks can be short or long-running; the protocol is built on HTTP, JSON-RPC, and server-sent events rather than a new transport.

According to Google's A2A launch post, the protocol shipped with support and contributions from more than 50 technology partners, including names such as Atlassian, Box, Salesforce, SAP, and ServiceNow. A2A launched with 50-plus technology partners.

MCP's job is narrower and older. Anthropic open-sourced MCP on November 25, 2024 as a USB-C-style port between models and data. The current MCP documentation still uses that USB-C metaphor: one protocol instead of a custom connector per system. According to the Linux Foundation's AAIF formation release, MCP had more than 10,000 published servers covering developer tools through Fortune 500 deployments when the foundation launched. MCP had 10,000-plus published servers at AAIF launch.

Anthropic's donation note adds two more adoption figures from the same week: Claude's connector directory listed over 75 MCP-powered connectors, and official SDKs saw 97M+ monthly downloads across Python and TypeScript. According to Anthropic, those official SDKs see 97M+ monthly downloads across Python and TypeScript. The MCP GitHub organization hosts the spec plus language SDKs, and its servers repository is the public list of maintained examples.

For a shop that already pushes web forms into a CRM, this split is concrete. The form connector is an MCP-shaped problem — see the form-to-CRM automation tools roundup for the operator tools in that layer. The moment that captured lead has to be worked by a contractor's agent, or by a client's own bot, you have left MCP and entered A2A.

A 10-person agency already mapping those form-to-CRM steps in US Tech Automations can add an A2A card on the same run instead of rebuilding the intake path.

The rest of the AAIF stack

A2A did not walk into an empty house. AAIF's own join post lays the layers out as five jobs.

AGENTS.md is the instruction file. agents.md calls it a README for agents: a predictable Markdown file at the repo root (and nested in monorepos) so coding agents get build steps, test commands, and conventions without stuffing a human README. According to OpenAI, AGENTS.md has been adopted by more than 60,000 open-source projects and by agent frameworks including Amp, Codex, Cursor, Devin, Factory, Gemini CLI, GitHub Copilot, Jules, and VS Code. AGENTS.md sits in 60,000-plus projects.

goose is the local-first runtime. Block contributed it at founding. The project now lives under AAIF stewardship; the docs at goose-docs.ai describe a desktop app, CLI, and API, with 45k+ GitHub stars, 500+ contributors, 70+ MCP extensions, and 15+ model providers. The code repo is github.com/block/goose (now served from the AAIF-goose org). A two-truck shop is not going to compile goose from source. The relevant fact is that the runtime the foundation hosts already speaks MCP, so a future A2A handoff does not require a second agent product just to reach files and APIs.

agentgateway is the traffic box. The agentgateway site describes one binary for HTTP, gRPC, LLM provider traffic, MCP, and A2A, and it states that the project has joined the Agentic AI Foundation. Install docs on that page ship Helm chart version v1.4.1. If you run agents on Kubernetes, this is the layer that can sign and scope tool calls and keep an audit trail of what an agent invoked.

The AAIF GitHub org currently lists seven working groups (identity and trust; governance, risk, and regulatory; workflows and process integration; accuracy and reliability; security and privacy; agentic commerce; observability and traceability) plus a taxonomy workstream. That is the committee layer sitting above the specs. It is not a product you buy.

Sources: Anthropic MCP introduction; Linux Foundation AAIF release; OpenAI AAIF note; goose docs; AAIF A2A join post; agentgateway.

Who shipped A2A, and who hosts it now

Google designed A2A from internal multi-agent work and launched it as an open protocol. The A2A GitHub repository still carries that history: Apache License 2.0, first commit dated April 9, 2025. The protocol homepage says A2A is maintained by a Technical Steering Committee with representatives from AWS, Cisco, Google, IBM Research, Microsoft, Salesforce, SAP, and ServiceNow.

On June 23, 2025, Google Cloud announced it was donating the specification, SDKs, and tooling to a new Linux Foundation Agent2Agent project, with AWS, Cisco, Google, Microsoft, Salesforce, SAP, and ServiceNow as founding members of that project. According to Google Cloud's donation post, more than 100 companies supported the protocol as of that date. That is the step Forbes is careful to separate from the August 2026 AAIF move: donation to Linux Foundation first, then a later shift into the Agentic AI Foundation alongside MCP.

IBM's Agent Communication Protocol merged into A2A in August 2025, which both Forbes and the AAIF join post record. Cisco's AGNTCY still covers overlapping discovery and identity ground, so "one protocol to rule them all" is not an accurate buyer slogan even after the IBM merge.

The Linux Foundation is the parent, not a startup. Its homepage states 1,300+ open source projects, 855K developers contributing code, and 21K+ contributing organizations. AAIF is one directed fund inside that structure. According to the Linux Foundation homepage, the organization hosts 1,300+ open source projects. That scale is why vendors were willing to put competing agent specs in the same building.

MCP's first-year community had already started meeting in public. The MCP anniversary spec post dated November 25, 2025 said the MCP Registry had close to two thousand entries and 407% growth from the initial batch onboarded that September. The MCP Dev Summit North America then ran April 2–3, 2026 in New York, with session recordings posted after the event. A2A arriving into that same foundation is a calendar fact, not a marketing merge of the two protocols.

Where the big clouds already list A2A

The AAIF join post is the primary list of production surfaces, and it names four.

Google Cloud supports developing and deploying A2A agents through ADK, Agent Engine, Cloud Run, and GKE, per AAIF. That matches the original launch posture on the Google Developers Blog.

Microsoft Azure AI Foundry lets agents expose A2A endpoints and find external agents through standard discovery, again per AAIF. According to Microsoft Learn, Foundry catalogs more than 10,000 models from Microsoft, OpenAI, Anthropic, Meta, and others. Foundry lists 10,000-plus models. That catalog figure is not an A2A conformance count; it is the size of the model shelf sitting next to Foundry's agent runtime. The Learn page was last updated on August 27, 2026.

AWS Bedrock AgentCore can host and operate A2A servers so agents built with different frameworks or hosted in different clouds can speak the same protocol, per AAIF. The AgentCore product page positions the service as a production platform for agents on any framework and any model. The AgentCore developer guide lists an A2A protocol contract beside HTTP, MCP, and AG-UI contracts, and it has a dedicated "Deploy A2A servers" path. Hosting an endpoint is not the same as implementing signed cards, multi-tenancy, and version negotiation; that gap is the one Forbes told enterprise buyers to probe.

Huawei standardized A2A between Celia, its OS-level assistant, and in-app agents on HarmonyOS, with Tencent's WeChat among the first major apps integrating over A2A for messages and calls under dual authorization, according to the AAIF join post. Google Cloud and PayPal are also extending A2A into agentic commerce through the Agent Payments Protocol (AP2), with shopping and merchant agents communicating over A2A and AP2 handling payment authorization, on that same page.

A two-truck HVAC shop that already sequences dispatch in US Tech Automations can keep that queue and publish a card so a parts vendor's agent can ask for a slot. The dispatch software question is the same one agencies already ask in the dispatch software for marketing agencies guide; A2A does not replace that board, it gives the board a way to speak to a vendor bot without a private API.

Sources: AAIF A2A join post; Google A2A launch; Microsoft Foundry overview; AgentCore developer guide.

USTA analysis: how long the stack took to sit in one house

USTA analysis. The dates below are taken only from the pages already cited. The lengths are calendar arithmetic on those dates, so a reader can redo the subtraction.

Inputs: A2A public launch on April 9, 2025; Linux Foundation donation on June 23, 2025; AAIF founding on December 9, 2025; A2A join on August 17, 2026. Partner floors: 50+ at launch, 100+ at donation, 150+ at AAIF join.

Sources: Google A2A launch; Google donation post; Linux Foundation AAIF release; AAIF A2A join post. Lengths are day counts on those dates; 3.0× is 150 ÷ 50 using the stated floors.

What the arithmetic is good for: it took about eight months after AAIF existed for the agent-to-agent layer to sit next to MCP, and about sixteen months after A2A's public launch. The partner floor tripled on the vendors' own stated numbers. That is not independent measurement — Forbes already warned that the 150 figure comes from the foundation and its vendors — but it is the only public count on offer, and the 3.0× multiple does not invent a new survey.

For a small shop, the useful reading is narrower. You did not miss a two-week window. The inter-agent rulebook spent a year in a Linux Foundation project of its own before it moved next to MCP. Buying software this quarter still means asking which revision you are getting, not assuming "AAIF" on a slide equals "our HVAC bot can call your parts bot tomorrow."

What the transfer does not solve

A shared foundation settles who stewards the text of the spec. It does not write the access policy for your tenant.

The A2A specification is explicit that agents stay opaque: they collaborate without sharing internal state, memory, or tools. A signed agent card authenticates the card's identity and metadata against a signing key. Forbes noted the limit that follows: that guarantee holds only as far as the recipient can establish that the key belongs to the organization being asserted. A forged or mis-issued key is still an enterprise identity problem, not a protocol problem.

Authorization is also left open on purpose. A2A requires a server to perform authorization checks and scope results to the caller's authorized boundaries, but the policy model itself is not a single enterprise ACL you can download. When a partner's agent is allowed to see last week's invoices, that decision still lives in your IdP, your gateway, and your audit log. agentgateway is one place to put that control plane; it is not installed for you when a vendor says "we support A2A."

Implementation depth varies. Hosting an A2A endpoint is cheaper to claim than supporting the v1.0 bindings, version negotiation, multi-tenancy, and signed cards. Consensus governance can also slow a release train; whether AAIF hosting changes A2A's cadence is not yet a measured fact.

Cisco's AGNTCY still overlaps discovery and identity, so a buyer can still be sold two "open" discovery stories. IBM's ACP merge removed one competing spec, not every competing spec.

None of that is a reason to ignore the move. It is a reason to put conformance, key provenance, tenant isolation, and auditability on the procurement checklist next to the logo slide.

Executive-assistant style work is a clean test case: calendar holds, inbox triage, and meeting briefs only stay safe if the delegated agent cannot see every mailbox. The executive assistant automation patterns still apply; A2A just makes the handoff to a specialist agent a protocol instead of a Zap.

Signal vs Speculation

Demonstrated fact, sourced above: AAIF formed on December 9, 2025 as a Linux Foundation directed fund. Founding projects were MCP, goose, and AGENTS.md. A2A launched April 9, 2025, was donated to a Linux Foundation project on June 23, 2025, reached spec 1.0.0, and joined AAIF on August 17, 2026. MCP, A2A, AGENTS.md, goose, and agentgateway are listed as AAIF projects. Google Cloud, Azure AI Foundry, AWS Bedrock AgentCore, and Huawei HarmonyOS are named as A2A surfaces by AAIF. The 50+, 100+, and 150+ partner floors are vendor- and foundation-stated, not third-party audits. Authorization policy and key provenance remain the operator's problem. MCP and A2A remain separate specifications.

Our read: if those partner floors hold and the three large U.S. clouds keep a documented A2A path, then over the next 12–36 months a small or mid-size business will meet A2A first as a checkbox on software it already buys — CRM copilots, dispatch boards, clinic intake — rather than as a project it staffs. The constraint that broke in 2026 was not model quality; it was the political cost of letting a competitor's protocol sit in a private Git repo. Putting A2A next to MCP lowers the chance that a shop is forced into one cloud's agent bus in order to talk to a partner.

Our read: the 12-month failure mode is "A2A-compatible" stickers on endpoints that do not negotiate versions or reject unsigned cards. The 36-month failure mode is a second discovery protocol winning a vertical (payments, telecom, a single mobile OS) and leaving SMBs to glue two "open" stacks. Neither is a prediction of prices, market share, or a date when a two-truck shop should rip out its current dispatcher.

Agencies comparing client-portal tools should keep the same skepticism they already apply to Plutio alternatives and marketing-agency workflow tools: a protocol badge is not a workflow. The badge is useful when the vendor can show a card, a revision, and a log.

What a small team should do this quarter

Start from the homepage if you need the catalog, then treat A2A as a procurement question, not a rewrite.

Ask every agent vendor four things in writing: which A2A revision (the spec current release is 1.0.0), which protocol bindings (JSON-RPC, gRPC, HTTP/REST), whether unsigned agent cards are rejected, and whether a cross-organization task leaves an audit record a person can read. If the vendor only says "we support A2A" with no revision, treat that as an endpoint claim.

If you already run document, intake, or dispatch automations, do not rebuild them to "be agentic." Keep the runs. Add a card when a partner actually has an agent that needs to call you. Teams already routing documents through US Tech Automations workflows will plug this in as a model swap, not a rebuild.

When you are ready to put those runs on a shared agent bus, open the agentic workflows platform and map one handoff — form to CRM, dispatch to parts, intake to billing — instead of a full stack replacement.

FAQs

What is the Agentic AI Foundation?

The Agentic AI Foundation is a Linux Foundation directed fund, announced December 9, 2025, that hosts open agent projects so no single vendor owns the rulebook. Founding contributions were Anthropic's MCP, Block's goose, and OpenAI's AGENTS.md, with A2A joining on August 17, 2026, as documented by the Linux Foundation and the AAIF join post.

Does A2A replace MCP?

No. The A2A project states that MCP is for agent-to-tool communication and A2A is for agent-to-agent communication, and that A2A is not a replacement for MCP. Use MCP to give one agent a calendar or a database; use A2A when that agent must delegate a task to another agent.

When did Agent2Agent join the Agentic AI Foundation?

AAIF announced the move on August 17, 2026, on its own blog. Google had already donated A2A to a Linux Foundation project on June 23, 2025, per the Google donation post.

Should a 10-person agency wait to buy A2A as a product?

There is no A2A SKU to wait for. A2A is a protocol vendors implement. Buy the workflow you need — intake, dispatch, billing — and require the vendor to name an A2A revision. If no partner agent will call you this year, you can keep MCP-style tool connections and postpone exposing a card.

What does a signed agent card actually prove?

It proves the card's identity and metadata verify against a signing key, as described for v1.0 in Forbes and in the A2A specification. It does not, by itself, prove the key belongs to the company named on the card, and it does not grant access to your data. You still bind keys to organizations and scope what a caller may see.

Who maintains A2A after the AAIF move?

The A2A homepage says a Technical Steering Committee with AWS, Cisco, Google, IBM Research, Microsoft, Salesforce, SAP, and ServiceNow maintains the project, under Apache License 2.0. AAIF hosting changes the foundation umbrella; it does not collapse A2A's maintainers into MCP's maintainers.

Glossary

Agentic AI Foundation (AAIF). A directed fund under the Linux Foundation that hosts open agent projects, including MCP, goose, AGENTS.md, agentgateway, and A2A.

Agent2Agent (A2A). An open protocol for one agent to discover another, send a task, and receive an artifact, without sharing internal memory or tools.

Model Context Protocol (MCP). An open protocol for connecting an AI application or agent to tools, files, APIs, and data sources.

Agent Card. A JSON metadata document an A2A server publishes so other agents can learn its identity, skills, endpoint, and auth requirements.

AGENTS.md. A Markdown file convention that gives coding agents project-specific instructions, now stewarded by AAIF.

goose. An open-source, local-first agent runtime contributed by Block, with MCP extensions, now under AAIF.

agentgateway. An open-source gateway for service, LLM, MCP, and A2A traffic, listed among AAIF projects.

Signed agent card. An A2A v1.0 feature that binds a card's identity and metadata to a cryptographic signing key; useful only if you can trust that key's owner.

The shared rulebook is now in one building. The work of making two agents interoperate still belongs to the teams who deploy them. If you want to put a single handoff on that rulebook, use the agentic workflow builder and start with one job, not a platform migration.