I've been working as a founding engineer at a startup and got frustrated with how fragile "Chain-of-Thought" agents are in production. They tend to hallucinate or drift off-topic over long horizons.
CyberLoop is an attempt to treat agentic reasoning as a control systems problem rather than just prompt engineering.
Key ideas:
Inner Loop (System 1): Cheap, deterministic probes (using code/grep/sensors) to explore without burning tokens.
Outer Loop (System 2): The LLM acts as the controller, only stepping in to plan or evaluate.
Thermodynamics: We treat "deviation from goal" as entropy and use "relaxation ladders" to dampen the oscillation.
The repo includes a reference implementation and a demo (GitHub search agent) that shows how it self-corrects when a search yields 0 results.
It's still a research preview (v1.0.1), but I'd love to hear your thoughts on applying control theory to agent orchestration.
roackb2•1h ago
I've been working as a founding engineer at a startup and got frustrated with how fragile "Chain-of-Thought" agents are in production. They tend to hallucinate or drift off-topic over long horizons.
CyberLoop is an attempt to treat agentic reasoning as a control systems problem rather than just prompt engineering.
Key ideas:
Inner Loop (System 1): Cheap, deterministic probes (using code/grep/sensors) to explore without burning tokens.
Outer Loop (System 2): The LLM acts as the controller, only stepping in to plan or evaluate.
Thermodynamics: We treat "deviation from goal" as entropy and use "relaxation ladders" to dampen the oscillation.
The repo includes a reference implementation and a demo (GitHub search agent) that shows how it self-corrects when a search yields 0 results.
It's still a research preview (v1.0.1), but I'd love to hear your thoughts on applying control theory to agent orchestration.
Link to paper: https://zenodo.org/records/17835680