Agent Interop & the Protocol Stack 2.0
One year on from MCP, models compressed to a ~3% tie and tokens down ~1,000x — so what's the moat? The rails. MCP / A2A / AG-UI are now neutral Linux Foundation infrastructure, and who owns the protocol layer is the $50B contest playing out in 2026.
goose by Block (aaif-goose)
goose (Block) has crossed ~46,700 GitHub stars with 500+ contributors and 110+ releases — a local-first, general-purpose agent you run as a desktop app, CLI or via API, with any LLM, wired to the MCP ecosystem (3,000+ servers and counting). What makes it the right spotlight: goose now lives under the aaif-goose organisation — one of the three founding projects of the Agentic AI Foundation (alongside MCP and AGENTS.md). It is the consumer-facing embodiment of the open protocol stack: MCP for tools, local-first for data residency, no SaaS lock-in, foundation-stewarded rather than vendor-owned. The whole thesis of this issue shipped as something you can install tonight. (OpenClaw still tops raw OSS charts at 210K+ as the local-first foil — viral, but not a governed protocol citizen.)
- ~46.7K stars — 500+ contributors, 110+ releases — local-first agent, any LLM, MCP-native
- 3,000+ MCP servers — the tool ecosystem goose plugs into via the open MCP standard
- 1 of 3 — AAIF founding projects (MCP + goose + AGENTS.md) — foundation-stewarded, not vendor-owned
1 · One year on — from experiment to infrastructure
Anthropic open-sourced the Model Context Protocol in late 2024 as a tidy answer to a messy problem: every agent needed bespoke glue for every tool. On its first anniversary (Nov 25, 2025) MCP shipped a major spec adding a Tasks API for long-running async jobs and Client ID Metadata Documents to simplify client registration and trust — upgrades you only build once a protocol is load-bearing. The adoption curve is the headline: from ~100K SDK downloads at launch to ~97M per month by early 2026 (~970x in 18 months), with Anthropic, OpenAI, Google, Microsoft, AWS and Salesforce all shipping support inside ~13 months, and 10,000+ public servers (59,000+ across all registries).
The decisive 2026 move was governance. On December 9, 2025 the Linux Foundation formed the Agentic AI Foundation (AAIF), anchored by three donated projects — Anthropic's MCP, Block's goose and OpenAI's AGENTS.md — with platinum members AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI. The protocols moved out from under any single vendor's wallet into neutral stewardship. Google's A2A, donated to the Linux Foundation back in June 2025, sits in the same umbrella. The MCP Registry is now funded by dozens of companies at $100K+/year — the 'protocol-win' playbook is now an institution, not a bet. The reframe: a year ago these were experiments competing for mindshare. Today MCP is plumbing — assumed, not argued.
2 · The three-layer stack — MCP / A2A / AG-UI
The 2026 production architecture is not one protocol but three, each answering a different question with a clean separation of concerns. You don't choose between them — you layer them. MCP connects an agent down to its tools and data; A2A connects agents sideways to each other across frameworks and orgs; AG-UI connects the agent up to the human in a live, streaming interface. AWS Bedrock AgentCore Runtime added AG-UI support in March 2026, cementing the three-layer stack as the reference production pattern.
The elegance is composition: an orchestrator uses A2A to hand a sub-task to a specialist agent, that agent uses MCP to hit the database and the payments API, and the whole run streams back to the user over AG-UI with approval gates and a kill switch. Get the layering right and you can swap the model underneath (Day 85's cheap, fungible token) without touching the wiring — which is why the wiring, not the model, is where durable value now sits.
| Protocol | Question it answers | Signature primitives | Steward |
|---|---|---|---|
| MCP (agent → tools) | How does an agent reach a tool, file, API or data source? | Tools / resources / prompts; Streamable HTTP + OAuth 2.1; Tasks API; MCP Apps (UI returns) | AAIF / Linux Foundation |
| A2A (agent ↔ agent) | How do agents discover, delegate to and trust each other across orgs? | Signed Agent Cards; task lifecycle; 5-language SDKs (Py/JS/Java/Go/.NET) | Linux Foundation |
| AG-UI (agent → user) | How does any front-end stay in sync with a running agent in real time? | 16 event types: token stream, tool calls, shared state, generative UI | CopilotKit / open spec |
3 · The interop frontier — where the rails are still being laid
Mature does not mean finished. The live work in 2026 is at the seams — making the three layers, and the packaging on top of them, genuinely interoperable rather than merely co-existing. Read the four frontiers together and a shape emerges: the protocol layer (MCP/A2A/AG-UI) is converging, and a packaging-and-discovery layer (SKILL.md + AGENTS.md + registries) is forming on top of it — the way npm sat on Node. That layer is where the App-Store economics of Day 60 play out, and it is being standardised now, while it is still free to plant a flag on.
| Frontier | What's shipping in 2026 | Why it matters |
|---|---|---|
| A2A ↔ MCP interop spec | Both are now Linux Foundation projects with overlapping membership and a committed joint interop spec | Formally bridges 'tool access' and 'agent-to-agent' so the seam stops being bespoke glue |
| MCP Apps | Tool calls return interactive UI (dashboards, widgets), not just text; stateless spec expected to land mid-2026 | Turns MCP from a data pipe into a UI surface — the bridge toward AG-UI |
| SKILL.md + AGENTS.md | Packaging layer on top of the protocols; SKILL.md spoken by 20+ agents, 8 marketplaces by Q2 2026 | The portable unit of agent value — write once, run on any compliant runtime |
| Registries | mcp.so (20K+), Smithery ('Docker Hub of MCP'), Glama, official MCP Registry — 59K+ servers indexed | Discovery infrastructure: an agent can find and wire a tool at runtime |
4 · Why the standards layer is the moat now
Stack the last two retrospectives. Day 80: frontier models are within ~3% of each other — capability is a near-tie. Day 85: the serving stack drove a GPT-4-class token down ~1,000x — the token is cheap and fungible. If neither the model nor the token is scarce, the scarce, defensible thing is the connective tissue that lets a cheap-token fleet actually work together, and the governance evidence that makes one agent safe to trust inside another's workflow.
Governance rides the rails. The same OTEL→WORM audit trail you owe the regulator (Day 81) and the buyer (Day 82), and the SPIFFE/SVID identity + kill switches from Day 54, are exactly what make your agent a trustworthy citizen of someone else's fleet. Interop without audit is a liability; interop with it is a sellable feature. 'Show me your Agent Card and your audit trail' is becoming the new integration question.
That is why the real 2026 contest is who owns the rails. Anthropic plays the open-protocol / npm game (donate MCP and Skills, win the ecosystem). Google plays the OS-surface game (Gemini, Android, Spark — own the device). Microsoft plays the enterprise control-plane game (Agent 365, Entra — own the identity and registry). All three routes lead to the same prize: whoever's rails the fleet runs on captures the next $50B even as model performance plateaus. And there is a clock on it: the asymmetry window — when the standards are open and free to build on — closes around 2027 as revenue shares formalise. The cheapest flag you'll ever plant in the agent economy is the one you plant this year.
With models within ~3% and tokens down ~1,000x, the moat has moved to the connective-tissue layer — the protocol stack and governance evidence that let a cheap-token fleet discover, call and trust each other. MCP/A2A/AG-UI are already neutral foundations; SKILL.md + AGENTS.md + registries are the packaging layer forming on top. Asymmetry window closes ~2027 as revenue shares formalise.
Use MCP for tool/data access, A2A for agent-to-agent delegation, and AG-UI for the user-facing surface. Adopting the stack means you can swap the model underneath (Day 85's cheap, fungible token) without rewiring — the wiring is the asset, so don't build a bespoke RPC you'll have to maintain alone.
Publish one SKILL.md to the marketplaces, expose MCP entry points on your product so agents can call it, and register an A2A Agent Card so other agents can discover it. The asymmetry window — free to build on, open governance — closes around 2027 as revenue shares formalise.
The same OTEL→WORM audit trail and SPIFFE/SVID identity you owe the regulator (Day 81) and buyer (Day 82) are what make your agent safe to trust inside someone else's fleet. Interop without an audit trail is a liability; interop with one is a feature you can sell.