AgentDecarbonizer Optimizes AI Agent Workflows for Lower Carbon Emissions

AI agents, which extend large language models into complex, goal-directed workflows, involve numerous model calls and interactions with external environments. These extensive operations can lead to significant carbon emissions. A new study characterized these emissions using OpenClaw agent workloads, revealing that token consumption, context cache reuse, and the carbon intensity of the electrical grid are key factors. The research identified that the flexibility in task deadlines presents a major opportunity for reducing carbon footprint. By strategically delaying tasks or shifting them to different geographical grids, agents can execute during periods when the energy grid's carbon intensity is lower. However, implementing such carbon-aware execution requires addressing challenges like uncertain task execution times for temporal shifting and the overhead of recomputing cached contexts during spatial shifting. To tackle these issues, AgentDecarbonizer was developed as a carbon optimizer for AI agents, designed to work alongside OpenClaw. Given a task prompt and a user-defined deadline, AgentDecarbonizer estimates task duration and selects deadline-feasible execution schedules. It specifically accounts for the cache recomputation overhead when considering spatial shifts. Evaluations on various agent tasks across multiple grids showed that AgentDecarbonizer can reduce carbon emissions by up to 57.9% compared to carbon-agnostic baselines and 37.5% against baselines that only select the optimal grid at task start.