AI is rapidly moving from single-response chatbots toward agentic systems, where Large Language Models (LLMs) plan, reason over multiple steps, use external tools, and self-correct to complete complex tasks. These systems are powerful, but they carry a serious and largely overlooked downside: efficiency. A single agent task can trigger dozens of LLM calls as the system plans, retries, reflects, and coordinates between sub-agents. So even when each individual call is modest, the energy and carbon cost of a full agentic workflow can be many times that of a simple one-shot query.
As agentic AI is adopted more widely, this hidden cost will become a real sustainability and affordability problem, particularly for smaller organisations, universities, and settings where compute, energy, and budgets are tight. This project proposes a framework for energy-efficient agentic AI built around a clear question: how can we design agentic LLM systems that complete tasks reliably while using far less computation and energy? Rather than building new models, the project focuses on the agent layer, where most of the waste actually happens, by making planning, tool use, model selection, and coordination far more efficient.
Aim
The aim of this project is to develop a framework for energy-efficient agentic AI that makes multi-step LLM agents significantly more sustainable, without sacrificing the reliability that makes them useful. The project will investigate where computational waste occurs in agentic workflows and design strategies to reduce it, including smarter planning, selective use of small versus large models, and the avoidance of unnecessary or repeated calls. Ultimately, the research aims to establish practical principles for building agentic systems that stay capable while operating within realistic computational and energy limits.
Objectives
This project will first investigate and measure where energy and computation are consumed within agentic LLM workflows, identifying the most wasteful patterns such as redundant calls, excessive retries, and over-reliance on large models for simple sub-tasks. Second, it will design efficiency strategies at the agent level, including adaptive planning that minimises unnecessary steps, and model-tiering that routes easy sub-tasks to small efficient models (SLMs) while reserving large models for genuinely hard steps. Third, it will develop mechanisms to reduce wasted computation, such as caching, early stopping, and confidence-based escalation. Fourth, it will build and evaluate complete agentic pipelines on representative tasks. Success will be measured across four areas: energy consumption, number and cost of model calls, task success rate, and overall reliability, allowing clear comparison between standard and energy-efficient agent designs.
Significance
Agentic AI is widely expected to be one of the defining directions of the next generation of intelligent systems, yet its sustainability cost is almost entirely unstudied. This project addresses that gap directly. By building energy-awareness into how agents plan, coordinate, and choose models, the research has the potential to dramatically reduce the energy and carbon cost of agentic workflows while keeping them effective. The project contributes new knowledge at the intersection of agentic AI, sustainable computing, and resource-efficient AI systems, and its outcomes are expected to support practical, environmentally responsible agents that can run affordably, including in edge and low-resource deployments. As agentic AI scales, efficiency at the agent layer may prove just as important as efficiency at the model layer.
Ideal Candidate
We are looking for a self-motivated PhD candidate with strong analytical, problem-solving, and research skills. Candidates with a background in Computer Science, Software Engineering, Artificial Intelligence, Data Science, or a related field are encouraged to apply. Experience with machine learning, deep learning, Python programming, and software development is desirable. Familiarity with large language models, agentic or multi-step LLM systems, natural language processing, model optimisation, or sustainable AI would be an advantage. Additionally, the applicants should meet the eligibility criteria for entry into a PhD program at Curtin University.
This project is open to International and Domestic applicants.
Internship
This PhD project may offer the opportunity for an industry internship as part of the candidature, providing valuable real-world research experience and exposure to applied AI in an operational setting. There is potential for the internship to be undertaken with an industry partner such as Thales or L3Harris, subject to availability and mutual interest. An internship of this kind would allow the candidate to apply their research on energy-efficient agentic AI to practical, resource-aware deployment challenges, while building industry connections and complementing their academic work. The specific scope, host organisation, and timing would be confirmed during the candidature.
Scholarship
If you are identified as the preferred candidate for this project, you may be considered for an RTP scholarship.
Enquires and How to Apply
For enquires about this opportunity contact Associate Professor Sajib Mistry at Sajib.Mistry@curtin.edu.au
To formally apply submit an Expression of Interest to Associate Professor Sajib Mistry during the Central Scholarship round (July 1st – July 31st 2026)