🧠 CyberLoop

The Thermodynamics of Intelligence

"Stop building open-loop agents. They drift. They fail. They burn money."

Figure 1: Conceptual model of stability in Agentic Systems. Standard agents (Red) tend to accumulate semantic entropy over time. CyberLoop (Blue) aims to dampen this oscillation through closed-loop feedback.

CyberLoop is the reference implementation of AICL (Artificial Intelligence Control Loop) — a framework that treats Agentic Reasoning not as a prompt engineering problem, but as a Control Theory problem.

Most agents today are Open Loop: they generate tokens based on probability, accumulating semantic entropy with every step. CyberLoop is Closed Loop: it introduces Deterministic Probes (Sensors) and Relaxation Ladders (Actuators) to enforce convergence towards a goal.

---

⚡️ Performance & Validation

CyberLoop is domain-agnostic. We validate its control mechanics across different problem spaces using the exact same control architecture.

1. Mechanism Demo: GitHub Repository Search

(Run this locally to see the control loop in action)

This demo illustrates how CyberLoop's Inner Loop handles exploration mechanics (narrowing/broadening) without burning LLM tokens on every step. While modern LLMs often memorize popular repos, this demo proves the self-correcting behavior of the framework when facing search constraints.

• Query: "node graceful shutdown"
• Mechanism: The agent uses Probes as gradient signals. When a search yields 0 hits, the probe fails, triggering a deterministic strategy switch (e.g., Narrow → Broaden) without LLM intervention.
• Run it: yarn examples:github

Step	Action	Logic
0	narrow	Initial strategy
...	...	...
5	narrow	Probe Failed (no-hits): Gradient signal detected
6	stop	Budget Exhausted: Bounded exploration triggered

2. Industrial Case Study: Root Cause Analysis (RCA)

(Internal Benchmark on OpenTelemetry Data)

To stress-test the framework in a high-dimensional production environment (where the answer isn't in the LLM's training set), we implemented a private RCA Adapter using CyberLoop to analyze distributed tracing data.

Compared to a standard Tool-Using Agent (GPT-5):

Metric	Standard Agent	CyberLoop (AICL)	Impact
LLM Calls	13 calls	2 calls	85% reduction (Solved inference cost)
Execution Time	109s	71s	34% faster (Deterministic inner loops)
Infinite Loops	Occasional	Zero	Idempotency detection
Convergence	Probabilistic	Mathematical	Guided by Ladder

Note: While the RCA domain adapter is proprietary, the underlying control logic is identical to the open-source GitHub demo.

---

🧬 Core Philosophy: Hierarchical Control

We replace "Prompting" with a System 1 / System 2 architecture:

1. Inner Loop (Reflexive System)

• Role: Fast, cheap, deterministic exploration.

• Cost: ~$0.01 per step.

• Mechanism: Uses Probes (Sensors) to detect if a path is viable, and Ladders to adjust exploration intensity.

• Key Constraint: Zero LLM calls allowed.

2. Outer Loop (Strategic System)

• Role: Slow, expensive, semantic planning.

• Cost: ~$0.50 per call.

• Mechanism: The LLM acts as the "Controller", setting the initial strategy and evaluating the final state found by the Inner Loop.

📖 Read the Deep Dive:

• The Whitepaper (AICL) - Theoretical foundation & architecture.
• Philosophy - Why Control Theory is the missing link for AGI.

---

🛠️ Getting Started

Installation

# Clone the repository
git clone https://github.com/roackb2/cyberloop.git
cd cyberloop

# Install dependencies
yarn install

Running the Demos

You will need an OPENAI_API_KEY and a GITHUB_TOKEN.

# 1. Run the AICL Agent (Closed Loop)
# Watch it broaden/narrow its search based on probe feedback
yarn examples:github

# 2. Run the Baseline Comparison
# Compare results with a standard "flat" agent
yarn examples:github:baseline

---

🏗️ Architecture & Modules

CyberLoop is built on modular, swappable interfaces.

Module	Role	Loop Layer
Environment	Provides observable states and executes actions	Both
ProbePolicy	Deterministic decision logic based on probe signals	Inner
Probe	Low-cost feasibility checks (e.g., "Is result set empty?")	Inner
Ladder	Regulates exploration entropy (Relaxation Gradient)	Inner
Planner	LLM-based strategy and final evaluation	Outer
BudgetTracker	Hard constraints on token/step usage	Both

See Inner/Outer Loop Architecture for the detailed sequence diagram.

---

📂 Documentation Index

• Theory: AICL Whitepaper | Philosophy
• Design: Engineering Guardrails | Trace Spec
• Applications: Use Cases Appendix (RCA, Code Search, Knowledge Retrieval)
• Academic: Zenodo Record (Cite as Liang, 2025)

---

Status: 🧩 Work in Progress (Research Preview)

Uncontrolled intelligence grows powerful but fragile. Controlled intelligence grows stable — and endures.

📜 Licensed under the Apache 2.0 License © 2025 Jay / Fienna Liang (roackb2@gmail.com)