H2A — Human to Agent: The Missing Piece
A2A connects agents. MCP connects tools. But who connects the human?
H2A · A2A · MCP · AG-UI · Modality Negotiation · NIST 800-63
A customer calls from their smartphone. The AI agent reads the appointment aloud and simultaneously sends a booking link via SMS. Same agent, same customer — now from a landline. No SMS possible. So the agent books entirely by voice. Then in the vehicle: two short sentences, yes-or-no question, a map on the display. The agent knows what its counterpart can do. That's H2A — Human to Agent.
Three Protocols, One Vacuum
Google built A2A for agent-to-agent: Cards, Tasks, Delegation. Anthropic developed MCP for agent-to-tool: Discovery, Schemas, Backend. CopilotKit filled the frontend gap with AG-UI: SSE Events, Shared State. But what happens when the agent doesn't know if SMS is possible? If the driver is distracted? If the display can render three buttons — or only voice?
The Capability Envelope
{
"channel": "headunit",
"static": {
"maxButtons": 3,
"markdown": false,
"driverDistraction": true
},
"situational": {
"driving": true,
"maxSentences": 2,
"yesNoOnly": true
}
}The solution: a structured JSON that each channel sends at connection setup. What am I? What can I do? What's happening right now? Two layers — static and situational. Static never changes within a session. Situational is updated live via events. The agent adapts in real time.
Static
- Head unit: 3 buttons, no markdown
- Webchat: rich content, images
- Landline: voice only, no SMS
- Never changes per session
Situational
- Vehicle driving → 2 sentences, yes/no
- Vehicle stopped → longer answers
- Map view possible when parked
- Live updates via events
One Agent, Every Touchpoint
The revolutionary part: the agent itself doesn't change. Same orchestrator, same specialist agents, same logic. Only response formatting adapts to the envelope. WhatsApp? New adapter, new envelope. Done.
Orchestrator
Coordinates all specialist agents via A2A. Discovers them automatically in the Agent Garden.
Identity Agent
Authentication and verification via AI Confidence Scoring.
Booking Agent
Appointment booking adapted to channel capabilities.
Vehicle Agent
Vehicle data and control via the One MCP Gateway.
Service Agent
Service requests with full CRM context.
One MCP Gateway
Central backend access: CRM, calendar, vehicle APIs, KB.
AI Confidence Scoring
Instead of rigid authentication: parallel signal fusion. Eleven methods, prioritized per NIST 800-63, regulatory secured per GDPR Art. 22. Voice biometrics is not a standalone factor — NIST explicitly prohibits that. FIDO2 and passkeys have priority.
The system is built on seven regulatory frameworks simultaneously: GDPR, eIDAS 2.0, PSD2, EU AI Act, NIST 800-63, ISO 21434, UNECE R155.
The Complete Protocol Stack
AG-UI
Frontend ↔ Agent: Events, State, Rendering. CopilotKit standard.
A2A
Agent ↔ Agent: Tasks, Delegation, Federation. Google standard.
MCP
Agent ↔ Tools: Discovery, Schemas, Backend. Anthropic standard.
SIP/WebSocket
Voice world: audio that no other protocol natively handles.
The Capability Envelope is the bridge that AG-UI lacks. It extends frontend-to-agent with modality negotiation, situational constraints, and auth context.
Why Nobody Builds This
Status Quo
- One bot per channel
- Each knows only its context
- Customer re-explains to each
- No cross-channel auth transfer
- Isolated islands
H2A Architecture
- One agent, all channels
- Envelope delivers context
- Seamless customer journey
- AI Confidence across channels
- Connected system
“Whoever skimps on the protocols between human and agent builds isolated islands instead of a connected system.”
No automotive company worldwide has prepared an EUDI-Wallet-Ready architecture. Nobody yet combines AI Confidence Scoring with regulatory first-principles design. That requires an architectural rethinking that goes beyond individual teams.
— Philipp
H2A for your organization? Let's talk about omnichannel AI architecture.
Book a consultation