Abstract
dc:description.abstractMulti-agent systems have become a dominant paradigm for tackling complex tasks in artificial intelligence, distributing work across networks of specialized agents that plan, reason, and act in coordination. As these systems grow in capability and scale, two important classes of challenges emerge, namely how to use agents to reliably manage complex computing infrastructure, and how to make the execution of multi-agent systems themselves more efficient. This thesis addresses both challenges. The first contribution is LADs, an agentic large language model based framework that automates the full configuration, deployment, and maintenance lifecycle of distributed cloud services. Misconfigurations are a leading cause of production failures in cloud environments, and many existing approaches address individual stages of the DevOps pipeline in isolation. LADs unifies these stages into a closed feedback loop that combines configuration generation with live deployment signals and iterative self repair. Evaluated across 180 deployment tasks spanning Dask, Redis, and Ray, LADs substantially improves deployment success rates, resolves the large majority of runtime failures within three iterations, and reduces cost by orders of magnitude compared to conventional developer driven workflows. The second contribution is a workload aware cache eviction policy for multi-agent systems. Many multi-agent systems decompose queries into execution plans represented as directed acyclic graphs, where nodes correspond to agent operations. Caching the results of these nodes can substantially reduce redundant computation across queries, but memory constraints necessitate eviction. Traditional eviction strategies treat cached entries uniformly based on access history alone, ignoring structural and workload signals uniquely available in agentic settings. Our policy incorporates recomputation cost, DAG dependency count, and agent invocation frequency into a unified scoring function, consistently outperforming all finite capacity baselines in latency and throughput across diverse benchmarks while approaching the performance of an unbounded cache.
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Mohamed, Anas
Rights
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Dc Identifier Citation
- Mohamed, Anas. (2026). Reliable and Efficient Multi-Agent Systems. Retrieved from the University Digital Conservancy, https://hdl.handle.net/11299/281682.