Communication Protocols: The Agentic DSL

When people collaborate, they use natural language full of politeness and context. When agents collaborate, politeness is a Financial Burden.

Example of Waste (Assistant to Specialist):

"Hello Research Agent. I hope you are well. Could you please find the price of gold today and report back to me in a JSON format? Thank you so much!" (45 tokens)

Example of Efficiency (The DSL):

REQ: gold_price | FMT: JSON (7 tokens)

In this lesson, we master the Agentic DSL (Domain Specific Language). We’ll learn how to build a strict, high-density protocol for inter-agent messaging that cuts communication costs by 80%.

---

1. The "Standard Header" Pattern

Just like HTTP has headers (Content-Type, Auth), your agentic system should have internal headers.

The Protocol:

• T: Task ID
• C: Context Level (1-5)
• I: Input Data
• O: Output Expectation

By sticking to this "Standard Header," you remove the need for the model to "Parse intent" from a sentence. The intent is fixed in the structure.

---

2. Using "Internal Call" Markers

In a multi-agent graph (LangGraph), you can use a specific symbol (like @) to trigger a specialist.

Agent to Specialist: @searcher: "Apple Stock 2024"

Specialist to Agent: RET: 150.22

This "Chatbot Shorthand" is perfectly understood by LLMs if they are instructed as "Technical Operators" in the system prompt.

---

3. Implementation: The Protocol Formatter (Python)

Python Code: Minifying Inter-Agent Messages

def format_inter_agent_msg(agent_name, task_body, output_format="text"):
    """
    Wraps a task in a token-dense protocol string.
    """
    protocol = f"[@{agent_name}]"
    protocol += f" B:{task_body}" # B for Body
    protocol += f" F:{output_format}" # F for Format
    
    return protocol

# Result: [@coder] B:fix_indentation_err F:diff
# VS: "Hey coder agent, please fix the indentation in this file and give me a git diff."

---

4. The "No-Preamble" Constraint

In your system prompt for every agent in the fleet, you must include a "Protocol Restriction."

The Preamble Constraint:

"INTERNAL COMMUNICATION ONLY. You are talking to another AI. DO NOT use greetings, signatures, or filler words. Output ONLY the protocol string. Failure to be concise costs money."

---

5. Token Savings: Scaling the Conversation

In a multi-agent swarm where 5 agents talk to each other to finish one task, "DSL Communication" can save thousands of tokens per user session.

Protocol	Avg Tokens/Msg	Total (Swarm of 20 Msgs)
Natural Language	60	1,200 tokens
DSL (Shorthand)	10	200 tokens
Savings	-	1,000 tokens

---

6. Summary and Key Takeaways

1. Delete Politeness: AI-to-AI talk should be technical, not conversational.
2. Define a Header: Use single-letter keys for task metadata (T, B, F).
3. Symbols over Words: Use @name to address specialists.
4. Enforce at System Level: Make conciseness a primary directive for all agents.

In the next lesson, Shared Context vs. Private Context, we look at چگونه to decide which agent knows what to prevent "Information Oversaturation."

---

Exercise: The Protocol Rewrite

1. Take a typical "Agent Instruction": "Dear SQL Agent, I need you to select the top 10 users from the database and send them back as a Markdown table."
2. Rewrite it in DSL format.
3. Validate: Ask an LLM to "Decode" your DSL.

• Does it correctly identify the Role, the Task, and the Format?
• (Usually, Yes).
• Count the savings. (It's usually ~75%).

---

Congratulations on completing Module 12 Lesson 3! You are now a protocol designer.