Where Should Agents Live? Energy-Memory Characterization of Agentic AI for the Edge-Cloud Continuum
Carolina Fortuna, Vid Han\v{z}el, Tim Strnad, Bla\v{z} Bertalani\v{c}
cs.AI · cs.LG · cs.MA
Quick verdict
Extends eCAL to multi-agent LLM energy via a validated two-rate model; text hand-offs cost <0.25% of workflow energy, while multi-agent setups rarely pay off.
Summary
This paper extends the eCAL AI-lifecycle energy metric to multi-agent LLM workflows (agentic-eCAL), combining a closed-form two-rate single-call model — compute-bound prefill linear in prompt tokens, memory-bandwidth-bound decode following a_dec/b + c_0 — with 7-layer OSI transport accounting and amortized pre-training energy per useful output bit. The call model is calibrated and validated on A100/vLLM with Qwen2.5-7B and Llama-3.1-8B (270 configurations each; R²>0.99, ~10% MAE) and cross-checked against published H100 numbers (Table II). Experiments over reasoning depth, team width (2–30 agents), and batch size show history-carrying loops scale super-linearly (exponent 1.02–1.42, rising with batch), and that KV-cache width, not weight class, bounds serving density across 16 open-weight models. Transmission analysis over 5G/metro/optical bearers finds inter-agent text costs <0.25% of workflow energy, while KV-cache migration (18.9 Gbit/session) is energetically non-transportable. On a 24-incident Kubernetes Infra-Bench edge benchmark, multi-agent topologies raise energy up to 23.9× (Qwen2.5-7B) and 4.0× (Qwen3.5-9B) per run without consistent remediation gains; energy per solved task reaches 5,222 kJ versus 49.7 kJ for the single-agent baseline. Main limitations: benchmark energy is inferred from token counts at an assumed batch, statistics are point estimates without variance, tool/retrieval terms are analytical, and code is not yet released.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- agentic-eCAL generalizes the eCAL per-bit metric to directed multi-agent workflows, and its two-rate single-call energy model matches NVML measurements with R²>0.99 and ~10% MAE on A100/vLLM.
- Inter-agent text transmission costs less than 0.25% of workflow energy across 5G RAN, metro, and optical bearers; the dominant cost of distribution is the additional prefill/inference induced by communication, not transport.
- History-carrying agentic loops scale super-linearly in energy (measured exponent 1.02-1.42, rising with serving batch) because accumulated transcripts are re-prefilled each step, and KV-cache width — not model weight class — is the binding constraint on serving density.
- KV-cache migration is energetically impractical: at 18.9 Gbit per session it exceeds compute energy on cellular bearers (parity intensity 5.1e-8 J/b by K=6), binding an agent's context to its local accelerator.
- On a 24-incident Kubernetes edge benchmark, multi-agent topologies consume up to 23.9× (Qwen2.5-7B) and 4.0× (Qwen3.5-9B) more energy than a single ReAct agent without consistently improving remediation success.
Strengths
Physically grounded and honestly validated call-energy model
Eqs. 4-6 decompose a call into compute-bound prefill and batch-amortized memory-bound decode; Fig. 2 validates on 270 A100/vLLM configurations per model (R²>0.99, MAE ~10%), recovers a_dec=3.2/3.3 J and c_0=0.02/0.06 J/token, and Table II transparently shows where the datasheet form under/over-predicts published H100 numbers instead of hiding the discrepancy.
Decision-relevant negative result overturning the classical offloading trade
Table III and Fig. 4d quantify inter-agent text transport at 0.003-0.186% of workflow energy across bearers (8.9 mJ to 0.52 J for 1,290 tokens), showing the dominant cost of distribution is the induced prefill on arrival, not transmission — directly actionable for operators deciding where agents should reside.
Novel hand-off protocol analysis with a clear feasibility boundary
Energy tied to delivered utility, not just runtime
Honest scaling characterization
Weaknesses
Headline benchmark energy is modeled, not metered
No uncertainty quantification on benchmark conclusions
Generalization beyond the validated regime
Analytical components carry argumentative load
Internal inconsistencies and unreleased code
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
How much energy does communication between AI agents actually cost?
According to this paper, very little: sending agent-to-agent text over 5G, metro, or optical links accounts for less than 0.25% of a multi-agent workflow's total energy (a 1,290-token hand-off costs between 8.9 mJ on metro fiber and 0.52 J on a loaded 5G cell). The real cost of distributing agents is the extra LLM inference triggered when a receiving agent re-reads the accumulated conversation, not the network transmission itself.
What is agentic-eCAL?
agentic-eCAL is a metric proposed in this paper that generalizes the earlier eCAL (Energy Cost of AI Lifecycle) metric to multi-agent LLM workflows. It sums LLM inference energy (split into a compute-bound prefill term and a memory-bandwidth-bound decode term), tool execution, vector retrieval, network transmission, and an amortized share of the model's pre-training energy, divided by the useful output bits produced — yielding joules per useful bit for an entire agent workflow rather than a single inference.
Do multi-agent LLM systems use more energy than a single agent?
On this paper's Kubernetes incident-repair benchmark (24 tasks), yes, substantially: multi-agent topologies consumed up to 23.9× more energy than a single ReAct agent with Qwen2.5-7B and 4.0× with Qwen3.5-9B, without consistently solving more incidents. Energy per successfully repaired task ranged from about 29-79 kJ for multi-agent setups versus 17 kJ for the single agent (Qwen3.5-9B), and up to 5,222 kJ versus 50 kJ for the smaller model.
Can you move an LLM's KV cache between edge and cloud servers?
The paper argues this is energetically impractical: a serialized KV cache is roughly five orders of magnitude larger than the equivalent text (18.9 Gbit vs 0.33 Mbit per session in their tests). At realistic link energy intensities, cache migration would exceed the compute energy it saves on cellular links, so an agent's context is effectively bound to the accelerator that built it and should be kept local rather than offloaded over the network.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Abstract is truncated mid-sentence and figure-embedded numbers (e.g., Fig. 7b, Fig. 9c) are partially garbled in the extracted text; figure panels were assessed via captions and surrounding prose only.
- The benchmark's serving hardware and the Qwen3.5-9B calibration procedure are not stated anywhere in the extracted text; the corresponding weakness is based on what the paper omits rather than on contradictory evidence.
- This is an arXiv v1 (16 Sep 2026) submitted to IEEE JSAC; several 2026-dated references and the Infra-Bench benchmark could not be verified offline.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.