Slipstream

Semantic Quantization for Multi-Agent AI Communication

---

82% fewer tokens. Semantic meaning preserved. Built for the AAIF ecosystem.

Before (45 tokens):
{"from": "alice", "to": "bob", "type": "request", "action": "review", "target": "auth_module"}

After (5 tokens):
SLIP v1 alice bob RequestReview auth_module

Multi-agent AI systems waste 40-60% of compute on coordination overhead. At scale, that's $180K-$2.5M/year just for agents talking to each other.

Slipstream fixes this through semantic quantization - transmitting pointers to concepts rather than the concepts themselves.

---

Quick Start

pip install slipcore

from slipcore import slip, decode, think_quantize_transmit

# Create a message (5 tokens instead of 45)
wire = slip("alice", "bob", "RequestReview", ["auth_module"])
# -> "SLIP v1 alice bob RequestReview auth_module"

# Or let the quantizer map natural language
wire = think_quantize_transmit(
    "Please check the authentication code for security issues",
    src="dev", dst="reviewer"
)
# -> "SLIP v1 dev reviewer RequestReview"

# Decode
msg = decode(wire)
print(msg.anchor.canonical)  # "Request review of work"

---

Why Slipstream?

The Problem

BPE tokenizers fragment compressed formats, negating syntactic optimization:

Compressed: REQ/TSK|s=7|d=3|act=review
Expected:   8 tokens
Actual:     22 tokens (every | and = is a token!)

The Solution

Slipstream uses a Universal Concept Reference (UCR) - a shared semantic manifold where common agent intents have single-token names that work across all LLM architectures.

Format	Tokens	Annual Cost (50 agents)
JSON verbose	~45	$180,000
JSON minimal	~30	$120,000
Slipstream	~5-8	$32,000

---

Wire Format

SLIP v1 <src> <dst> <anchor> [payload...]

• No special characters - avoids BPE fragmentation
• Space-separated - clean tokenization
• CamelCase anchors - single tokens in most tokenizers

Core Anchors

Category	Anchors
Requests	RequestTask, RequestReview, RequestHelp, RequestPlan
Inform	InformComplete, InformProgress, InformBlocked, InformStatus
Propose	ProposePlan, ProposeChange, ProposeAlternative
Evaluate	EvalApprove, EvalReject, EvalNeedsWork
Meta	Accept, Reject, MetaAck, MetaHandoff, Fallback

---

Finetuned Model

We provide a ready-to-use model trained on the Slipstream protocol:

Format	Link	Use Case
LoRA Adapter	slipstream-glm-z1-9b	Merge with base
GGUF Q4	slipstream-glm-z1-9b-gguf	Ollama / llama.cpp
Dataset	slipstream-tqt	Train your own

Run with Ollama

ollama run anthony-maio/slipstream

Train Your Own

# Generate training dataset
python -m slipcore.finetune_llm -n 1000 --provider gemini -o train.jsonl

# See notebooks/slipstream_finetune_colab.ipynb for full guide

---

AAIF Integration

Slipstream is designed as the transport layer for the Linux Foundation Agentic AI ecosystem:

┌─────────────────────────────────────┐
│   Application (Agent Logic)        │
└────────────────┬────────────────────┘
                 │
┌────────────────▼────────────────────┐
│   MCP / A2A (Semantic Layer)        │
└────────────────┬────────────────────┘
                 │
┌────────────────▼────────────────────┐
│   Slipstream (Transport Layer)      │  ← 82% token reduction
└────────────────┬────────────────────┘
                 │
┌────────────────▼────────────────────┐
│   Network                           │
└─────────────────────────────────────┘

---

Resources

• Paper: Slipstream: Semantic Quantization for Efficient Multi-Agent Coordination
• Model: HuggingFace
• Dataset: HuggingFace
• Spec: spec/slip-spec.md

---

Citation

@misc{maio2025slipstream,
  title={Slipstream: Semantic Quantization for Efficient Multi-Agent Coordination},
  author={Maio, Anthony},
  year={2025},
  url={https://github.com/anthony-maio/slipcore}
}

---

License

Apache 2.0

---

Stop paying the token tax.

pip install slipcore