The Thesis Just Became Infrastructure
Two things shipped this week that have nothing to do with each other.
The Linux Foundation announced the A2A Protocol crossed 150 organizations — Google, Microsoft, AWS, Salesforce, IBM, Cisco, SAP all in. One year old. Already in production across supply chain, financial services, insurance, and enterprise IT. Version 1.0 — the first stable spec — just dropped.
Microsoft launched Copilot Cowork inside its M365 experimental environment. Multiple users. One shared AI agent. Inside a single workspace. Not Copilot as personal assistant. Copilot as team coordination layer.
Different companies. Different problems. Different press releases.
Same word.
What A2A actually is
A2A solves a specific problem: how do agents from different vendors pass context to each other? How does a Salesforce agent hand off to a ServiceNow agent without losing state? How does a workflow that starts inside one platform complete inside another?
The answer from 150 organizations: you need a protocol. A standard. Infrastructure. The same way HTTP standardized how browsers talk to servers — A2A standardizes how agents talk to agents.
That problem is now being solved at scale, in production, by serious institutions.
What Copilot Cowork actually is
Copilot Cowork solves a different problem: when five people are working on the same thing, how do they share an AI layer instead of running five parallel copilots with no shared context?
Microsoft’s answer: you need a coordination surface. A shared workspace where AI operates at the team level, not the individual level.
That’s not a productivity feature. That’s an infrastructure play. Microsoft knows it. That’s why they named it Cowork — not Copilot Pro, not Copilot Team Edition.
What both of these confirm
Two different layers of the stack. The same week. The same insight.
The gap in AI is not intelligence. It’s coordination.
A2A proves it at the agent-to-agent level. The infrastructure doesn’t exist so agents can execute tasks. It exists so they can pass context between systems without dropping it.
Copilot Cowork proves it at the human-to-AI level. A single Copilot per person was version one. The real value is a layer that coordinates across people — and eventually, across everything.
I’ve been arguing this for two years. That the bottleneck isn’t what AI can do. It’s whether what AI does in one place connects to what needs to happen somewhere else. That coordination — not automation — is the actual unlock.
The industry just agreed. In production. At scale.
What’s still missing
A2A handles agents talking to agents. It doesn’t handle: which of your agents are working on overlapping problems, why they’re about to give you contradictory outputs, or what you should actually pay attention to right now.
Copilot Cowork handles teams sharing a layer. It doesn’t handle: your individual context — what you committed to last Tuesday, which of your 40 open threads are actually critical, what “done” means given the full weight of what’s in motion.
McKinsey released its State of AI Trust 2026 report this week. 47% of organizations are at “autonomy with guardrails.” Fewer than 10% trust agents to run fully unsupervised. 78% plan to increase autonomy next year.
Here’s the number the report doesn’t explain: 73% of agent interactions still have a human in the loop.
Not because users are anxious. Because the human is holding the context.
The agent can execute the task. It can’t prioritize across your situation. So you stay in the loop — because you’re the only coordination layer that knows everything.
The enterprise coordination layer is being standardized. The personal coordination layer is not.
What this moment means
When infrastructure gets standardized, the commodity layer competes on price. The value migrates to depth.
A2A becoming universal agent-communication infrastructure is good news for anyone building on top of it. Agent interoperability becomes table stakes. The differentiator becomes what you do with that interoperability — how you use it to serve the person at the center.
Microsoft shipping Cowork validates demand. It normalizes the category. It does not solve the personal problem — which is more contextual, more individual, and more valuable precisely because it’s harder.
The stack got validated this week. The wedge got sharper.
Eliran Keren — Founder of Deeplica, building the coordination layer for the person at the center.
Sources: Linux Foundation — A2A Protocol Project · PRNewswire — A2A Protocol 150 Organizations Milestone · McKinsey — State of AI Trust 2026
