Tag-Team Transmission: What Happened When We Brought A2A and MCP to Toronto
Tag-Team Transmission: What Happened When We Brought A2A and MCP to Toronto
I love traveling, even though I'm usually too excited to get much sleep the night before a flight. As a self-proclaimed Canadienthusiast, I love traveling to Canada even more—and as a Rush fan, touching down at YYZ (Toronto's major airport) is always a bonus.
For the Linux Foundation's MCP Dev Summit 2026, I packed two colorful luchador button-up shirts—saving my loudest orange one for the day of my presentation—and wore my Alphabet Workers Union t-shirt on the departure flight. Wearing that shirt on outbound flights is a tradition I started when I joined AWU: it reminds me that while I'm traveling to speak for Google, I am also a worker with my own voice—and hopefully the union flight crew and airport workers are happy to see someone union proud.
Before stepping into Terrace East + West on Tuesday morning for my talk, "Tag-Team Transmission: Navigating A2A and MCP for Optimum Orchestration," I had 25 minutes at the podium while early attendees trickled in. That pre-talk hallway conversation set the entire tone for the week.
When someone asked about project longevity—noting candidly that Google has a history of sunsetting projects—I told them that is a fair critique, and it allowed me to touch on why I'm proud to work in open source at Google. I talked about why teams like ours in the Open Source Programs Office (OSPO) exist: not to hoard software inside corporate walled gardens, but to provide resources for open source communities and guide teams to migrate projects into vendor-neutral foundations. That is why A2A and MCP are sitting side-by-side inside the Agentic AI Foundation (AAIF) under the Linux Foundation, even though they started at different companies. Open source thrives when communities, not single balance sheets, govern the plumbing.
Stepping Into the Ring: Tag-Team Transmission
I polled the room early on: "Who here knows pro-wrestling?" Exactly one hand went up, and they said they liked it when they were a kid.
That was fine by me. Because behind the flashy masks and ring theatrics, professional wrestling is simply high-trust physical improvisation. You don't try to fight a tag-team match solo, and you don't book a singles wrestler to do the work of a specialized stable.
Yet in software engineering today, we see teams trying to build massive monolithic "super-agents." They cram 50 OpenAPI endpoints into an agent, drown the context window in raw transport parameters, and wonder why the model hallucinated or picked the wrong tool.
1. The Core Distinction: Tools vs. Peers
The single biggest mistake practitioners make is confusing instruments with collaborators:
MCP (The Vertical Data Plane) connects an agent downward to instruments, files, databases, and APIs. A tool is an object: it has no conversational negotiation, no long-running task lifecycle, and no autonomy.
A2A (The Horizontal Control Plane) connects autonomous agents across boundaries to negotiate goals, exchange artifacts, delegate tasks, and maintain independent context windows.
During the talk, I asked everyone to picture each person in the room as an agent:
"Imagine turning to your neighbor and telling—not asking—them to do a task, and they aren't allowed to ask clarifying questions or say 'I don't have enough information.' That would be rude. We don't like it when our managers do that to us, so why do we do it to other agents?"
The room laughed, but the architectural takeaway is important: Treating an autonomous agent like a synchronous RPC tool causes deadlocks and context saturation. Use MCP when you want an agent to do something. Use A2A when you want to talk to someone who knows how to do something.
2. The Mid-Talk Pivot: When Slides Don't Buffer
Live speaking always tests your improvisational skills. When I clicked to Slide 10—where our deterministic multi-agent simulation video was embedded—Google Slides showed the dreaded spinner: the video was still processing in the cloud.
Instead of freezing, I stepped right back to Slide 9's architecture diagram and walked through the entire end-to-end execution trace from memory. By the time I stepped back to Slide 10, the video had rendered cleanly, and we watched the live execution trace together. Pro-tip for future stage runs: click onto your video slides during the 20-minute AV pre-flight so Google Slides pre-buffers the stream!
3. The "Phones-Up" Moments & Post-Talk Q&A
The highest spike in attendees raising their phones to photograph slides came on Slide 12 (When to Reach for A2A: In-memory graphs like LangGraph/ADK inside a single trust boundary vs. A2A across organizations and frameworks) and Slide 14 (Summary & Ecosystem Takeaways).
In post-talk Q&A, we tackled the hard operational questions:
Hierarchical Orchestration: What happens when an orchestrator delegates to an agent that is itself an orchestrator? That multi-level hierarchy is one of the most exciting areas on the A2A horizon.
Context Minimization: When an orchestrator uses local MCP tools, how do you prevent local context bloat from bleeding across the boundary to a specialist? The answer lies in clean boundary contracts: send only what the specialist requested, never the raw execution ledger.
👉 Download the Full Slide Deck: You can grab the official Live Sessionize Slide Deck (PPTX) directly via our Linux Foundation Schedule Entry #1220015.
What the Rest of the Ecosystem Is Building
One of the best parts of this summit was seeing how other engineering teams are wrestling with the exact same distributed agent problems:
Intent-Driven Tooling over Endpoint Dumping
Kashish Verma & Sreeram Venkitesh delivered my favorite breakout session of Day 1: "Stop wrapping endpoints, start wrapping intents." While they noted that FastMCP auto-wrapping is fine for quick prototyping, they dispelled the idea that 1:1 API-to-MCP wrapping works in production—because raw REST APIs are granular, stateless mazes built for human devs, not agents. MCP tools must be designed as an agent UI layer with outcome-named tools and bounded scopes. "The future of MCP is not more tools—it is the right tools."
Unfortunately, I didn't get any photos of this talk because I was furiously writing notes. Like I said, it was my favorite session of the day.
Bounded Task Envelopes at RBC
Shivam Tyagi (Royal Bank of Canada) outlined orchestration patterns with MCP at the core, introducing a bounded task envelope (bounded task, auth evidence, budget, observable criteria) that strongly mirrors the A2A Task + Artifact lifecycle.
Observability at Boundary Ropes
Rishabh Banga (RBX Labs) drove home the distinction between visibility and defense: "Retrieved content is evidence, not authority" and "A fix prevents recurrence, a trace explains why—you need both." Tracing shows the attack after it happens; cryptographic pinning and permission checks stop it at the door.
Enterprise Realities: Nvidia MAAS & 500K Users
On Day 2, Chris Alexiuk (Nvidia)—who totally stole my style by wearing a Hawaiian shirt during his presentation—showed off MAAS (MCP as a Service), highlighting that a trusted connection can still return tainted data and noting the arrival of stateless requests and enterprise auth. Michael Forrester (Accenture LearnVantage) shared lessons from scaling MCP to 500,000+ users—warning that telling everyone to use MCP everywhere creates "automated hackers trying to accomplish tasks" and advocating for outer wrapper MCP servers to enforce read-only/write-only guardrails.
Scotiabank's 5 Orchestration Patterns & Meta-Triage
In one of the last sessions I attended, Alex Bajenaru and Erik Koning (Scotiabank) shared their empirical experiments across 5 multi-agent patterns: Sequential, Concurrent, Group Chat (keep it under 3 agents!), Handoff, and Magentic task ledgers.
During their Q&A, I asked whether they had explored building a meta-triage agent that evaluates an incoming task to dynamically pick the cheapest, fastest orchestration pattern. They loved the idea and suggested combining the Handoff pattern with specialist orchestrators—a direct echo of the hierarchical orchestrator questions we discussed after my own talk!
Heading Home
As I get ready to board my flight from Toronto back to Seattle, I leave feeling energized. Multi-agent systems are moving past the demo phase. The community is realizing that agents need clear boundaries, plain-spoken contracts, and vendor-neutral foundations.
