🤖 machine learning

AgentDecarbonizer: Carbon-Aware Execution for AI Agents

This paper introduces AgentDecarbonizer, a carbon-aware execution optimizer for AI agents that reduces carbon emissions by up to 57.9% by leveraging deadline flexibility to schedule tasks during low-carbon-intensity periods or on cleaner grids while accounting for execution uncertainties and cache recomputation overhead.

Original authors: Leyi Yan, Shuangning Li, Sihang Liu

Published 2026-08-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Leyi Yan, Shuangning Li, Sihang Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Every time a computer performs a calculation, it draws electricity from the power grid. That electricity is not always clean; depending on where the power comes from and when it is generated, the process can release varying amounts of carbon dioxide into the atmosphere. In recent years, a new kind of software has emerged that acts like a digital assistant capable of complex, multi-step thinking. Unlike a simple chatbot that answers a single question and stops, these autonomous agents can plan a long journey, break it into smaller tasks, use tools to gather information, and correct their own mistakes along the way. They can write code, analyze data, or manage schedules without constant human supervision. However, because these agents run for longer periods and make many more requests to the underlying artificial intelligence models than a standard chat, they consume significantly more energy. As these tools become more common, the carbon footprint of their endless thinking is becoming a serious environmental concern.

Researchers at the University of Waterloo and the University of Chicago have developed a new system designed to make these digital assistants more eco-friendly without slowing them down or making them less effective. They call their tool AgentDecarbonizer. To understand how it works, one must first look at how these agents operate. When an agent is given a task, it does not just think once; it thinks in a loop. It asks a question, receives an answer, uses a tool like a web search or a file editor, and then uses that new information to ask the next question. This cycle repeats many times. The researchers found that this process is incredibly heavy on "input," meaning the agent sends a vast amount of text to the computer model to keep it informed of the conversation history and previous steps. Because the agent often repeats parts of this history, the computer system can save energy by remembering what it has already processed, a technique known as context caching. If the agent stays in the same place, it can reuse these saved memories. If it moves to a different location, it must forget and relearn everything, which wastes energy.

The team discovered that the carbon cost of running these agents depends heavily on two things: where the computer is located and when the work is done. The electricity grid in different parts of the world relies on different energy sources. Some regions get most of their power from wind, solar, or hydroelectric dams, which produce very little carbon. Others rely more on fossil fuels like coal or gas, which produce much more. Furthermore, the mix of energy changes throughout the day. A grid might be very clean at noon when the sun is shining but become dirtier at night when solar power is unavailable. The researchers observed that an agent running in one region at a specific time could produce nearly three times the carbon emissions of the same agent running in a different region or at a different time.

To solve this, the researchers built AgentDecarbonizer to act as a smart scheduler that runs alongside the agent. Instead of letting the agent start its work immediately, the system looks ahead. It estimates how long the task will take and checks the forecast for electricity cleanliness in different parts of the world. If the agent has a flexible deadline, such as a task that needs to be finished within twenty-four hours, the system can wait for a time when the grid is cleanest. It can also decide to move the work to a region where the air is cleaner. However, the system is careful not to move the work just because a region looks cleaner on paper. It calculates whether the energy saved by using clean power would be wasted by the effort of moving the agent's memory to a new location. If the cost of moving and relearning is too high, the system keeps the agent where it is, even if the local power is slightly dirtier.

The researchers tested their system using a collection of sixty realistic tasks, ranging from analyzing scientific papers to debugging computer code. They ran these tasks across four different power grids with varying energy mixes. The results showed that by carefully choosing when and where to run the work, the system could cut carbon emissions dramatically. When given a full day to complete a task, the system reduced emissions by up to 57.9 percent compared to a standard approach that ignores carbon. Even when the deadline was tight, with only three hours to finish, the system still managed to reduce emissions by over 34 percent. The system proved that it is possible to make these powerful digital assistants more sustainable by simply being smarter about when and where they do their work, ensuring that the future of artificial intelligence does not come at the cost of the planet's climate.

Technical Summary: AgentDecarbonizer: Carbon-Aware Execution for AI Agents

Problem Statement

Autonomous AI agents extend Large Language Models (LLMs) from single prompt-response interactions into long-running, goal-directed workflows involving repeated model calls, tool invocations, and environmental interactions. While these workflows enable complex tasks like software repair and data analysis, their computational intensity leads to substantial carbon emissions. Existing carbon-aware optimization techniques, which typically shift workloads across time and geography to exploit lower carbon intensity, fail to address two unique characteristics of AI agents:

Uncertain Execution Time: Agent workflows are iterative and dynamic; later steps depend on previous outputs, making execution duration difficult to predict. This complicates the "temporal shifting" strategy, as agents must meet user-specified deadlines rather than deferring indefinitely.
Context Caching Overhead: Agents heavily reuse context (input tokens) across multiple model invocations. Modern LLM serving systems cache these contexts to avoid recomputation. Spatial shifting (moving execution to a different grid) often invalidates these caches, forcing the recomputation of previously cached inputs. This recomputation overhead can negate the carbon benefits of moving to a lower-carbon grid.

Methodology

The authors propose AgentDecarbonizer, a carbon optimizer designed to run alongside the OpenClaw open-source agent system. The system operates by jointly optimizing temporal and spatial execution decisions while respecting deadlines and cache locality.

1. System Architecture

AgentDecarbonizer consists of four main components (Figure 7):

Carbon Intensity Predictor: Forecasts future carbon intensity for candidate grids using historical data and energy mix measurements (utilizing the EnsembleCI predictor).
Execution Time Estimator: A lightweight local LLM (Gemma 4 4B) that estimates task duration. It performs a conservative initial estimate based on task phases and complexity, then refines the remaining time estimate in-flight using real-time progress, token counts, and API call statistics from the OpenClaw agent.
Cache-Aware Carbon Model: A planner that models context caching as a location-dependent penalty. It calculates the carbon cost of recomputing cached inputs if execution shifts to a new grid. A spatial shift is only selected if the expected carbon reduction from the new grid's lower intensity exceeds the recomputation overhead.
Execution Controller: Manages the runtime flow, pausing execution during "wait" intervals, resuming from saved states, and routing requests to specific grids based on the planner's schedule.

2. Optimization Algorithm

The core optimization problem is formulated as minimizing predicted carbon emissions plus cache-recomputation overhead, subject to deadline feasibility.

Formulation: The objective minimizes $\sum \hat{E}_{r,t} \cdot \widehat{CI}_{r,t} + \sum \Lambda^{(k)}$ , where $\Lambda^{(k)}$ represents cache-recomputation penalties incurred when switching regions between execution intervals.
Dynamic Programming: Under the assumption that cache recomputation depends only on whether the service region changes between consecutive intervals (local cache recomputation), the authors derive an exact dynamic programming solution. This algorithm runs in $O(K|R|H)$ time, where $K$ is the number of execution intervals, $|R|$ is the number of regions, and $H$ is the number of candidate time slots.
Deadline Guarantee: The system ensures deadlines are met by using conservative duration estimates. If the estimated remaining time exceeds the time left until the deadline, the system falls back to immediate execution on the current grid.

3. Characterization of Workloads

The authors characterized 60 AI agent tasks from WildClawBench across six categories (Productivity, Code, Social, Search, Creative, Safety). Key findings include:

Input-Heavy: Input tokens constitute 99.1% of processed tokens, with 94.6% being cached.
High Cache Savings: Context caching reduces carbon emissions by an average of 80.7%.
Variability: Execution times and API call counts vary significantly across tasks, necessitating dynamic estimation rather than static assumptions.

Key Contributions

Systematic Characterization: The first work to systematically study the carbon emissions of AI agents, identifying that emissions are driven by token consumption, context cache reuse, and grid carbon intensity.
AgentDecarbonizer Design: A novel carbon optimizer that handles the uncertainty of agent execution times and the overhead of context cache recomputation during spatial shifting.
Cache-Aware Optimization: A dynamic programming approach that explicitly balances the carbon benefits of grid shifting against the penalty of losing cached context.
Evaluation: A comprehensive evaluation demonstrating that the system meets deadlines while significantly reducing emissions.

Results

The system was evaluated using WildClawBench workloads (60 tasks) across four grids (Austria, California, Finland, Great Britain) with deadlines ranging from 3 to 24 hours. The model API used was GPT-5.4.

Carbon Reduction:
- Compared to a carbon-agnostic baseline (average emissions across grids), AgentDecarbonizer reduced emissions by up to 57.9% (with a 24-hour deadline) and 34.9% (with a 3-hour deadline).
- Compared to a current-optimal baseline (selecting the lowest-carbon grid only at the start time), it reduced emissions by up to 37.5% (24-hour deadline) and 3.3% (3-hour deadline).
Deadline Adherence: In all evaluations, the system successfully completed workloads within the user-specified deadlines.
Overhead: The local execution overhead on a Mac mini (M4) was minimal, with an average power consumption of 14.9 W and an execution time overhead of 0.7–5.2 minutes per workload.
Ablation Studies:
- Multi-grid Migration: Shifting across grids contributed an additional 34.1–45.7% reduction in emissions compared to waiting-only strategies.
- Prediction Robustness: The system remained robust to carbon intensity prediction errors, outperforming static one-time predictions, especially for longer deadlines.
- Estimator Accuracy: The conservative execution time estimator ensured that actual runtimes were consistently shorter than estimated times, guaranteeing deadline compliance.

Significance and Claims

The paper claims to be the first work to systematically study and optimize the carbon emissions of AI agents by tailoring methods to their specific execution characteristics. The authors emphasize that while autonomous agents amplify the computational demands of LLMs, their unique features—specifically the reliance on context caching and the variability of execution time—require specialized optimization strategies that differ from traditional carbon-aware computing.

The significance of AgentDecarbonizer lies in its ability to reconcile the conflicting goals of carbon reduction and deadline adherence in a setting where spatial shifting is not trivially beneficial due to cache recomputation costs. By demonstrating that carbon emissions can be reduced by up to 57.9% without violating user deadlines, the work provides a practical pathway for deploying sustainable, large-scale AI agent systems. The authors note that while the implementation targets OpenClaw, the optimization strategies are broadly applicable to other agent systems.