New Research · Frugal AI
Consolidation over Collaboration: The Frugal AI Case Against Multi-Agent Sprawl
A new research paper from Intics.ai shows why Multi-Agent Systems quietly inflate token costs and energy draw on edge hardware — and how a Multi-Skill Single Agent (MSSA) architecture keeps execution in the efficient, near-linear zone. Get the full paper, free.
Inside the paper
A rigorous look at why token cost, not raw capability, decides which agent architecture survives contact with edge hardware. Inside, you'll find:
The Multi-Skill Single Agent (MSSA) model
A single-agent architecture that consolidates skills instead of routing them across a network of sub-agents.
The math behind the context avalanche
A formal model showing why context ingestion grows quadratically with reasoning steps in any agentic loop.
Where MAS quietly inflates cost
How coordination turns and duplicated system prompts push Multi-Agent Systems deep into the expensive zone of the curve.
AFNOR SPEC 2314 alignment
How the MSSA architecture maps to the Frugal AI standard for measuring and reducing AI's environmental footprint.
A controlled benchmark on Intel edge silicon
Naive MAS, optimized MAS, and MSSA compared head-to-head on a 45W Intel Core Ultra edge workstation.
New evaluation metrics
The Token Efficiency Ratio (TER) and Joules per Task, proposed as standardized benchmarks for agent efficiency.
KV-cache persistence for skill switching
How runtime-level caching lets a single agent pivot between tasks without re-evaluating baseline context.
Why parallelism is an illusion on edge hardware
Why MAS's parallel-processing advantage disappears under real memory and power constraints — and what that means for deployment.
Written by the team building it
This paper is sponsored by Intics.ai, an Intelligent Document Processing (IDP) company, and authored by Janarthan Poornavel (CTO), Anil Kumar Sannareddy (CEO), and Vivek Acharya (CPO). A few questions we hear most often:
What is a Multi-Skill Single Agent (MSSA)?
An architecture that consolidates multiple operational skills inside one LLM boundary, instead of routing tasks across separate specialized agents. It removes inter-agent orchestration entirely.
Why do Multi-Agent Systems (MAS) cost more in practice?
MAS frameworks replicate system prompts and document context across every agent, and spend additional tokens on coordination and handoffs. The paper models this as a quadratic cost driver that MAS pushes into sooner and harder than a single agent does.
What hardware was this tested on?
An Intel Core Ultra (Series 2) H-Series processor at a 45W TDP with 64GB of shared LPDDR5x memory, running an INT8-quantized 8B model via the Intel Distribution of OpenVINO Toolkit.
What is the Token Efficiency Ratio (TER)?
A proposed benchmark measuring successfully completed tasks per million tokens consumed — a standardized way to compare agent architectures on cost-efficiency rather than raw capability.
How does this relate to Frugal AI and AFNOR SPEC 2314?
AFNOR SPEC 2314 is a French standard for measuring and reducing the environmental impact of AI systems. The paper frames MSSA as a practical way to meet that standard on constrained edge hardware.
Who is this paper for?
Engineering and product leaders evaluating agentic architectures for edge or cost-constrained deployments, especially in document-heavy workflows like IDP.
Turn efficiency research into an edge deployment plan
See how Intics.ai applies the MSSA architecture to Intelligent Document Processing workloads — and what it could mean for your own token costs and hardware footprint. To Download the Paper, Please help us with your information.