@@PROTOCOL — Token-Efficient LLM Commands

85%

Fewer Tokens

Stream

Parseable

Zero

Escaping

Why @@PROTOCOL?

❌ JSON-RPC (Standard)

{
  "tool": "write",
  "params": {
    "path": "src/app.ts",
    "content": "import express from 'express';\nconst app = express();\napp.listen(3000);"
  }
}

Tokens: ~60
Escaping: Required (\n, \")
Failure: One missing } = Total crash

✅ @@PROTOCOL (At-At)

@@WRITE_FILE src/app.ts
import express from 'express';
const app = express();
app.listen(3000);
@@END_FILE

Tokens: ~8
Escaping: Not required (raw content)
Failure: Graceful partial execution

Full Command Surface

Command	Usage	Description
`@@READ_FILE`	`@@READ_FILE /path/to/file`	Read file contents
`@@WRITE_FILE`	`@@WRITE_FILE /path ...content... @@END_FILE`	Write or overwrite file
`@@MKDIR`	`@@MKDIR /path/to/dir`	Create directory (with parents)
`@@RUN_CMD`	`@@RUN_CMD npm install`	Execute shell command
`@@DISPATCH`	`@@DISPATCH crew-coder Write auth endpoint`	Send task to specialist agent
`@@PIPELINE`	`@@PIPELINE [{"wave":1,...}]`	Multi-agent orchestration

Why It Works

🎯 High-Signal Markers

The @@ symbol is visually distinct and rarely appears in natural text, making it easy for LLMs to learn and reliable for parsers to detect.

⚡ Stream-Parseable

Unlike JSON (which must close before parsing), @@ commands are parsed line-by-line during streaming — reducing latency and enabling graceful partial execution.

📦 No Escaping Required

LLMs don't need to escape newlines, quotes, or special characters inside @@WRITE_FILE blocks. This eliminates common hallucination patterns where models forget to close strings.

🛡️ Graceful Failure

If an LLM hits a context limit mid-generation, every completed @@ block executes successfully. With JSON-RPC, one missing brace invalidates the entire payload.

🧠 LLM-Native

LLMs think in top-down narratives. @@PROTOCOL follows this: agents can explain why, then execute the @@ command, then continue their narrative — all in one natural flow.

💰 Token Efficient

JSON syntax (braces, quotes, escapes) consumes 15-30% more tokens than necessary. @@PROTOCOL reduces overhead from ~60 tokens to ~8 tokens per file operation.

Protocol Comparison

Feature	JSON-RPC	@@PROTOCOL
Token Overhead	45-60 tokens per call	8 tokens per call
Escaping	Required (`\n`, `\"`)	Not required
Streaming	Must close block before parsing	Line-by-line parseable
Partial Execution	One error = total failure	Graceful partial execution
LLM Hallucination	Frequent (unclosed braces)	Rare (simple markers)
Readability	Verbose nested JSON	Clean, narrative flow

Technical Specifications

v1.0 — Inline Syntax

The current production protocol uses inline @@ markers that agents embed in their natural language replies.

View v1.0 Specification →

v2.0 — Universal Adapters (Draft)

The upcoming v2 adds adapter layers for MCP, JSON-RPC, and OpenAI-style tool calls while keeping @@ syntax as the canonical internal format.

View v2.0 Draft →

Reference Implementation

• Orchestration: crew-lead.mjs
• Parsing: lib/dispatch/parsers.mjs
• Execution: gateway-bridge.mjs

Ready to deploy Special Agent Double @@?

@@PROTOCOL is built into crewswarm. Clone the repo and start building with token-efficient agent commands today.

Download crewswarm Read Full Spec