Designing Lightweight AI Agents for Edge Deployment
A Minimal Capability Framework with Insights from Literature Synthesis
Abstract
This thesis introduces the Minimal Capability Design (MCD) framework, a systematic methodology for engineering lightweight AI agents specifically optimized for edge deployment environments. Unlike conventional approaches that add memory, orchestration, and redundancy by default, MCD begins with architectural minimalism as a foundational design principle.
Through extensive simulation testing and domain-specific walkthroughs, this research demonstrates that stateless, prompt-driven agents can achieve robust task performance while maintaining interpretability and deployment efficiency under strict resource constraints.
Key Contributions
- Theoretical Framework: Formal design principles treating minimalism, statelessness, and prompt resilience as primary constraints
- Empirical Validation: Browser-based simulation testbed with quantized models (Q1/Q4/Q8) demonstrating MCD effectiveness
- Practical Application: Domain-specific walkthroughs in appointment booking, spatial navigation, and failure diagnostics
- Diagnostic Methodology: Systematic tools for detecting over-engineering and capability excess in AI agents
🧱 Part I: Foundations
Establishes the motivation for lightweight agents, reviews literature across architectural domains, and outlines the research methodology.
🧱 Part II: MCD Framework
Introduces core MCD principles, layered architecture, and practical instantiation strategies for stateless agent design.
🧩 Part III: Validation
Comprehensive testing through simulations, domain walkthroughs, comparative evaluation, and future research directions.
Downloads
PDF, EPUB, .JSON
🔗 Sharable Online Repo links
Online Links