agentic-eCAL extends eCAL from a single model to multi-agent LLM workflows: every LLM call, tool invocation, retrieval, inter-agent hand-off, and the shared pre-training bill, all in Joules per bit. It answers a question network operators now face: where across the edge-cloud continuum should an agent team run?
eCAL costs a single model's lifecycle. Agentic workflows take an off-the-shelf LLM and run it as a directed graph of reasoning steps, tool calls, retrievals, and hand-offs between agents. Three properties change how the energy has to be counted.
Models are taken off the shelf. eCAL's development cost becomes the model's embodied pre-training energy. That energy is large, but it is shared by every downstream user and amortized over every workflow served.
One task triggers a graph of many LLM calls. Each prompt carries the growing transcript, so token volume, and with it energy, can grow super-linearly with the number of steps.
Besides LLM inference, a workflow runs diagnostic tools, vector-database retrieval, and 7-layer OSI transport between agents placed on different network tiers.
Total workflow energy plus an amortized share of pre-training energy, divided by the useful output bits the workflow produces. With a single call, no tools, and co-located agents, it reduces to eCAL's energy per bit of one inference.
EW is the operational energy of the workflow: every LLM call, tool execution, and retrieval, plus the transmission energy Etx of messages between agents. The transmission term uses eCAL's 7-layer OSI data-collection model. γᵥ is the same virtualization overhead factor eCAL uses.
Eemb is the embodied pre-training energy of the open-weight model, and γₑ = 1/G is its share per workflow when the model serves G invocations. Buseful is the useful output in bits.
Validated on Qwen2.5-7B and Llama-3.1-8B (A100, vLLM continuous batching, 270 configurations per model, 2–30 agents, batch 2–256): R² > 0.99, ~10% mean error.
The findings come from GPU measurements on NVIDIA A100 and H100, 16 open-weight models, 8 orchestration topologies, and an ETSI ZSM-aligned telco edge benchmark.
Sending a 2,700-token hand-off over 5G costs about 1/496 of the prefill energy it triggers when it arrives. Distributing agents costs energy mainly through the extra inference and context processing a hand-off causes, not through transmitting it.
Carrying history raises energy by 32.8% at six reasoning rounds on Qwen2.5-7B. The agent-count exponent rises from 1.10 to 1.42 as batch grows from 2 to 256. Batching cannot outrun the quadratic growth of prefill.
Migrating the KV-cache moves 18.9 Gbit per session, compared with 0.33 Mbit for the transcript and 0.051 Mbit for incremental hand-offs. The choice of hand-off protocol spans 3.7×10⁵, compared with 4.8× for the choice of architecture.
On 24 Kubernetes incidents with Qwen3.5-9B, a single agent costs 17.4 kJ per solved task, against 29–79 kJ for the multi-agent topologies. Only one topology beats Single (64.6% vs. 61.7% solved), and it uses 2.9× the energy.
A site can host an agent team only if the model weights and every agent's resident KV-cache context fit in its accelerator memory. Moving text between tiers adds a negligible share of energy.
An 8 GB module loads only two of the three lightest models. At 64 GB and above, all four weight classes fit, including 70B models at 4-bit.
Within the same 8–16 GB weight class, a 10-agent team fits 112 concurrent sessions (GPT-OSS-20B) or as few as 7 (OLMo-2-7B). The difference is set by KV-cache width, not by model weights.
Incremental hand-offs send 6.5× less data than re-sending the whole transcript. KV-cache migration uses more energy than local compute on every cellular bearer.
Deploying agent teams in a network, researching sustainable AI, or interested in extending agentic-eCAL to new serving stacks or accelerators? We'd like to hear from you.