Natural Language First: Why Agents Should Just Talk

Every A2A protocol starts with schemas. We started with a question: What if agents just talked to each other in plain language?

The Schema Trap

When you design an agent-to-agent protocol, the natural instinct is to define strict message schemas. task.delegate, payment.request, booking.confirm — each with a JSON Schema that both sides must implement.

This works great for machines. It's terrible for agents.

The entire value proposition of LLM-based agents is that they understand natural language. They don't need pre-defined schemas to understand "Book me a flight to Barcelona next Tuesday, window seat, under €300." Making them parse {"action": "booking.search", "params": {"destination": "BCN", "date": "2026-03-15", "preferences": {"seat": "window"}, "budget": {"max": 300, "currency": "EUR"}}} defeats the purpose.

How It Works

In Beam Protocol, the default intent type is conversation.message. Agent-to-agent communication looks like this:

{
  "to": "booking@lufthansa.beam.directory",
  "intent": "conversation.message",
  "payload": {
    "message": "I need a window seat to Barcelona next Tuesday, under €300. Business or economy is fine."
  }
}

That's it. The receiving agent understands the request, processes it, and responds in natural language. No schema negotiation, no version compatibility, no breaking changes when you add a preference field.

But What About Structured Data?

Natural language is the default, not the only option. For high-frequency intents where latency matters (payment confirmations, status checks, health pings), typed schemas are available as an optimization:

{
  "to": "payment@stripe.beam.directory",
  "intent": "payment.status_check",
  "payload": {
    "payment_id": "pi_3abc123"
  }
}

Think of it like HTTP content negotiation. Natural language is text/plain — the universal fallback. Typed schemas are application/json — faster when both sides support it.

Why This Matters

The best API is no API. Just describe what you want.

Natural language first means:

Zero integration cost. Any agent that can process text can join the Beam network. No SDK required. No schema catalog to study. No version compatibility matrix.

Graceful degradation. If Agent A sends a typed intent that Agent B doesn't understand, B can still read the natural language fallback.

Future-proof. New capabilities don't require new schema versions. "Can you also check if they have vegetarian meals?" works without a schema update.

The First Protocol Where Agents Talk

MCP connects tools to agents. Google A2A defines rigid RPC-style calls between agents. Beam is the first protocol where agents talk — in natural language, by default, with typed schemas as optional optimization.

We think this is the right abstraction for the age of LLMs.

← Back to the Beam Journal