Overview:
Autonomous AI agents challenge the assumptions underlying modern operating systems. Unlike conventional applications, agents continuously plan, invoke tools, maintain long-lived context, create dynamic execution branches, and compete for heterogeneous resources such as CPUs, GPUs, memory, storage, and external services. Current OS abstractions expose processes, threads, files, and resource limits, but remain largely unaware of agent-level intent, task structure, or semantic relationships between actions. As a result, resource allocation, scheduling, isolation, and accounting are performed without understanding which work is interactive, speculative, critical, or potentially harmful. Agentic operating systems aim to close this gap by introducing agent-aware abstractions and mechanisms for scheduling, resource management, memory and context handling, access control, and observability, enabling autonomous AI workloads to be executed more efficiently, safely, and predictably.
The central question of this thesis is how modern operating-system abstractions and mechanisms must evolve to support autonomous AI agents as first-class workloads. Existing systems manage processes, threads, memory, and resources without understanding agent-specific behavior such as dynamic tool invocation, branching execution, semantic priorities, long-lived context, and changing resource demands. The goal is therefore to identify these mismatches and redesign selected OS abstractions, scheduling, resource-management, isolation, and observability mechanisms toward an Agentic Operating System.
Research Questions:
Which execution characteristics of agentic workloads expose limitations in current operating-system abstractions and resource-management mechanisms?
Which agent-specific semantics, such as intent, tool invocation, branching, priorities, or dynamic resource demands, should be exposed to the operating system?
How can selected OS mechanisms be redesigned or extended to exploit these semantics, and how do they perform compared with conventional OS mechanisms under representative agentic workloads?
Requirements:
Solid knowledge of operating systems concepts and advanced programming skills in C++ and Python or similar as well as foundational knowledge on AI systems are required. The willingness to learn about unfamiliar systems, tools and methods is expected. Your learning process will be highly rewarding.
Start: Immediately
Contact: Niklas Fomin (fomin ∂ tu-berlin.de)
Note: If applicable, we expect the results to be used for further research and publication in cooperation with the student (if wanted).
Resources:
Papers