Your AI agent has a carbon footprint. Not the datacenter's footprint, not the training run's — your agent's, for the specific task it just ran. And nobody is measuring it.
I just published a white paper on this (open access, DOI 10.5281/zenodo.23282699). Here's the plain-language version.
Every carbon tool we have stops at the boundary of a single model call. Cloud dashboards give you monthly totals by service. EcoLogits estimates per-API-call energy. All useful — and all blind to what an agent actually is.
A production agent task isn't a query. It's a workflow: a planner call on a frontier model, retrieval and tool calls, code execution, retries when something fails, a verifier pass at the end. Each step can run on a different model, a different provider, in a different region, on a different grid. No per-call meter sees the whole thing.
How much energy does one AI query use? There is no single answer — and that's the point:
That's a span of more than 1,000x. "A query" is not a unit of physics. And agents are the multiplier: a typical 4-6 call agentic task runs about 1.1 Wh, or ~50 Wh with reasoning enabled. The unit we should be accounting is the task, not the call.
The paper proposes a per-task carbon ledger. The math is simple on purpose:
CO2e per task = sum over steps of (energy x PUE x grid carbon intensity) + amortized embodied carbon
Every figure gets labeled honestly: measured, modeled, disclosed, or illustrative. No black boxes. The ledger also allocates the shared stuff no per-call meter sees — idle reserve capacity, KV-cache effects across steps.
Take a ~50 Wh reasoning task. Run it on the UK grid (233 gCO2e/kWh, PUE 1.54): about 18 gCO2e. Route the same task to Quebec's hydro grid (31 g/kWh): about 2.4 gCO2e. Same task, same answer — 7x less carbon, from routing alone. That's not a measurement, it's arithmetic from published inputs, but it shows why the ledger matters: you can't optimize what you can't see.
Here's the part I'm most interested in. We already budget dollars per agent task. The paper argues for a carbon budget alongside it: the orchestrator chooses model x region x timing per step to stay under a CO2e budget, the same way it stays under a dollar budget. Research prototypes doing this for single queries have cut modeled emissions ~51% — nobody's doing it across a whole agentic workflow yet.
Two deadlines are converging. The EU AI Act already requires model providers to document energy use (fines up to 15M euros) — but admits no measurement standard exists yet. California's SB 253 forces large companies to report Scope 3 emissions from 2027, and purchased AI inference lands squarely in Scope 3. Somebody has to define the methodology. This paper is my bid for what it should look like.
It builds on the Green Software Foundation's vocabulary — they name "per workflow execution" as the agentic functional unit — and extends it into something you can actually implement.
The full paper is open access: DOI 10.5281/zenodo.23282699. Read it, argue with it. And one question: if your agent platform showed the carbon cost of every run next to the dollar cost — would it change how you build?