"Stop building open-loop agents. They drift. They fail. They burn money."
Figure 1: Conceptual model of stability. Standard agents (Red) accumulate entropy over time.
CyberLoop (Blue) dampens this oscillation through closed-loop feedback.
"Your agent doesn't need a bigger brain. It needs a better body."
CyberLoop is the reference implementation of AICL (Artificial Intelligence Control Loop). It treats Agentic Reasoning not as a prompt engineering problem, but as a Control Theory problem.
In v2.1, we introduce Semantic Kinematics: giving agents a "vestibular system" (inner ear) to detect drift in vector space without needing to "think" (query an LLM). This allows agents to navigate complex knowledge graphs using pure mathematics—making them 100x faster and cheaper than traditional Chain-of-Thought agents.
Can an agent find the link between "Coffee" and the "French Revolution" without an LLM?
Using CyberLoop v2.1, the agent navigates Wikipedia using only Embeddings + PID Control. It "senses" semantic proximity and "reflexively" corrects course when it drifts or gets bored.
# 1. Install dependencies
yarn install
# 2. Run the Deep-Dive Scenario (Coffee -> French Revolution)
yarn examples:wikipedia revolution
| Metric | Pure LLM Agent (Typical) | CyberLoop v2.1 (Actual) | Impact |
|---|---|---|---|
| Decision Mechanism | Reasoning (LLM) | Sensing (Vector + PID) | Physiological |
| Latency per Step | ~3,000ms | ~50ms | 60x Faster |
| Cost per Step | ~$0.01 | ~$0.0001 | 99% Cheaper |
| Behavior | Stochastic | Controlled | Reproducible |
🔗 See full benchmark: docs/benchmarks/wikipedia/benchmark-wikipedia-navigation.md
CyberLoop separates the "Thinking" (Outer Loop) from the "Moving" (Inner Loop).
The reflexive system that handles fast, deterministic navigation and exploration. Zero LLM calls.
- 🛡️ Probe (Sensor): Low-cost feasibility checks. It asks simple questions like "Is the result set empty?" or "Did the API return 404?" to generate immediate feedback signals.
- 🪜 Ladder (Actuator): Mechanisms to regulate exploration entropy. If a probe fails, the ladder adjusts parameters (e.g., relaxing search filters, expanding candidate pools) to overcome friction.
- 🧭 Kinematics Engine (v2.1): Uses EKF/PID controllers to detect "Semantic Drift". If the agent's path diverges too far from the goal vector, the engine applies a "Correction Force" (Backtracking).
- ⚡️ Reflexes & Guards: Middleware for state hygiene. Includes Line-of-Sight (immediate action) and Boredom Penalty (avoiding loops).
The strategic system that handles planning, replanning, and final evaluation.
- Role: Semantic planning and high-level judgment.
- Mechanism: LLM (GPT-4o / Claude 3.5).
- Trigger: Only activated when the Inner Loop budget is exhausted or a stable state is found.
- Cost: Expensive, but rarely invoked.
Before v2.1, we validated the control loop concepts in industrial settings.
(Internal Benchmark on OpenTelemetry Data)
We used CyberLoop to analyze distributed tracing data in a production environment.
| Metric | Standard Agent | CyberLoop (AICL) | Impact |
|---|---|---|---|
| LLM Calls | 13 calls | 2 calls | 85% reduction |
| Execution Time | 109s | 71s | 34% faster |
| Infinite Loops | Occasional | Zero | Idempotency detection |
git clone https://github.com/roackb2/cyberloop.git
cd cyberloop
yarn install
Create a .env file with your keys:
OPENAI_API_KEY=sk-...
# v2.1: Wikipedia Deep-Dive (Semantic Kinematics)
# Scenarios: 'tech' (Jacquard -> CPU) or 'revolution' (Coffee -> French Revolution)
yarn examples:wikipedia revolution
# v1.0: GitHub Search (Deterministic State Machine)
# A classic example of bounded exploration using Probes & Ladders
yarn examples:github
- Benchmarks: Wikipedia Navigation Results
- Theory: AICL Whitepaper
- Architecture: Inner/Outer Loop Spec
- Academic (v2.1): Zenodo Record - The Brain Needs a Body (Liang, 2026)
- Academic (v1.0): Zenodo Record - AICL Whitepaper (Liang, 2025)
Status: 🧪 v2.1 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)