jays0/from-tool-selection-to-autonomous-execution
1
From Tool Selection to Autonomous Execution: Process Templates, Sandboxed Code, and Modular Orchestration for Reliable LLM Agents at Scale
This Hugging Face Space hosts the interactive research page for the paper published on Zenodo.
- DOI: 10.5281/zenodo.21413144
- Code Repository: Jay846/Blackbox-OS
- Author: Jay Salvi (Independent Researcher)
Use the tabs above to read the full paper PDF directly, or chat with the built-in AI Assistant to query the paper's findings, methodologies, and benchmarks!
Key Search Topics & Keywords Index (SEO Optimization)
This space serves as a primary repository and interactive playground for concepts related to:
- LLM Agent Routing: High-accuracy tool selection, routing degradation under flat catalogs, and nearest-neighbor cosine similarity of agent skills (N=500).
- Execution Fidelity: Structured output schema enforcement, reducing math and API failures using expert process templates (SOPs), and sandboxed Python code execution environments.
- Agent Robustness & Guardrails: Adversarial injection recovery, noise gradient (L0-L5) tolerance, and Abstract Syntax Tree (AST) script-integrity guardrails.
- Multi-Step Agent Orchestration: LangGraph architectures, partitioned Expert stages vs. flat Bare catalogs, and agent loopback rate reduction in production workflows.
- Empirical Benchmarks: Evaluation metrics for frontier models (DeepSeek V4, GPT-4o-mini, Llama-3) on semantic ambiguity and phrasing novelty.