The energy price of planning : energy-performance trade-offs in LLM-based software engineering agents
Abubakar, Abdulkadir (2026)
Diplomityö
Abubakar, Abdulkadir
2026
School of Engineering Science, Laskennallinen tekniikka
Kaikki oikeudet pidätetään.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi-fe20260819118446
https://urn.fi/URN:NBN:fi-fe20260819118446
Tiivistelmä
Software teams now hand routine coding work to agents built on large language models. To make these agents more reliable, developers add planning on top of the model: extra reasoning steps, self-critique, a memory of past tasks, or several attempts to choose from. These tactics often raise success rates, but they all work the same way, by making more model calls. More calls mean more time and more energy on every run, and that energy cost is rarely measured.
This thesis measures it. We tested five planning methods across three sizes of the Qwen3 model on a local code-editing agent, about 1,500 runs in all, recording the energy of each complete run together with whether it succeeded.
The results are clear. Heavier planning and bigger models both spent more energy without solving more tasks. The strongest setup was the simplest method on the mid-sized 8B model: it solved the most tasks and used the least energy per solved task. A method that worked at one model size could fail at another. And once we counted the runs that drew energy but returned nothing, the setup that looked most efficient proved the worst. Planning method and model size should be chosen together, and judged on energy as much as success.
This thesis measures it. We tested five planning methods across three sizes of the Qwen3 model on a local code-editing agent, about 1,500 runs in all, recording the energy of each complete run together with whether it succeeded.
The results are clear. Heavier planning and bigger models both spent more energy without solving more tasks. The strongest setup was the simplest method on the mid-sized 8B model: it solved the most tasks and used the least energy per solved task. A method that worked at one model size could fail at another. And once we counted the runs that drew energy but returned nothing, the setup that looked most efficient proved the worst. Planning method and model size should be chosen together, and judged on energy as much as success.
