Master's thesis: Tracing Agentic AI for Efficient and Secure Execution

Master's thesis: Tracing Agentic AI for Efficient and Secure Execution

The Connected Intelligence Unit is part of RISE Computer Science in Kista. The current research focus is on devising intelligent autonomous systems for controlling and allocating resources in future computer and communication networks. Among the group's key technologies are the Internet of Things (IoT) and Edge computing. The unit conducts projects together with industry and academic partners from Sweden and across the world.

Background and Purpose AI applications are increasingly evolving from individual LLM requests into agentic systems, where AI agents interact with external tools, software environments, data sources, and other agents to complete complex tasks. These applications can exhibit very different execution patterns. Some are sequential, while others involve parallel tool calls, multiple interacting agents, and complex dependencies between individual steps. As a result, agentic applications are fundamentally different from traditional chat-based interactions.

However, we still have a limited understanding of these workloads from a systems perspective. Existing traces and characterization tools often capture only part of the execution, for example model requests or high-level tool invocations, while providing little information about the underlying computational resources, data movement, dependencies, and system interactions generated by the complete application.

A better understanding of these workloads is important for designing future AI infrastructures. Detailed workload traces could, for example, help decide how different parts of an agentic application should be scheduled across CPUs, GPUs, network accelerators, and other computing resources. They could also help identify security-relevant behavior, such as access to sensitive data or resources, interactions with untrusted tools, and isolation requirements between different workloads.

Thesis Description The goal of this thesis is to develop methods and tools for detailed tracing and characterization of agentic AI workloads. You will study existing agentic applications, workload traces, and characterization approaches, and investigate what information is currently missing to accurately describe their execution.

Based on this analysis, you will design and implement an enhanced tracing framework able to capture both the high-level structure of an agent workflow and its interaction with the underlying system. Relevant information may include execution dependencies, concurrency, CPU and memory usage, storage and network activity, data movement, and interactions between models and external tools.

Using the collected traces, you will characterize different classes of agentic applications and investigate whether patterns can be identified across workloads. The resulting characterization can then be used to explore how agentic workloads could be executed more efficiently and securely. Possible directions include identifying opportunities for acceleration or offloading, and studying security and isolation requirements associated with agent and tool execution.

Terms: Start Time: As soon as possible

Scope: 30 hp

Location: RISE Computer Science, Kista, Stockholm. Option to partially work remotely.

Who are you? We expect you to have good programming skills, especially in Python, and an interest in computer systems and AI. Knowledge of Linux, distributed systems, performance profiling, and machine learning are important. You should enjoy experimental systems work and exploring emerging technologies where many research questions are still open.

We are looking forward to receiving your application! To know more, please contact Mariano Scazzariello ([email protected]). Applications should include CV, recent grades, and a code excerpt. Candidates are encouraged to send in their application as soon as possible but at the latest by the 1st of November 2026. Suitable applicants will be interviewed as soon as applications are received.