Master thesis Context Engineering Strategies for Agentic Development

Problem statement

As AI agents handle increasingly complex long-running development tasks a critical challenge has emerged: managing limited context windows across multiple agent continuous development scenarioswhere agents work on the same codebase over days or weeksagents must maintain coherence recall previous decisions and avoid redundant work within strict token constraints.

Current approaches often focus on isolated sessions or rely on pre-computed retrieval (RAG). However optimal performance requires thoughtful strategies across the entire agent lifecycle: pre-session preparation intra-session dynamic retrieval and post-session persistence. At Bosch Lund where we work extensively with multi-agent systems understanding which context management strategies provide the best balance of continuity performance and efficiency is crucial for production-ready agentic systems.

Proposed solution

This thesis investigates context management techniques across all three lifecycle stages in continuous development scenarios. The goal is to evaluate and compare different approaches allowing flexibility and discovery throughout the research.

Pre-session strategies: Initialization approaches (project docs previous summaries) selective vs. comprehensive loading preparation overhead vs. effectiveness trade-offs.

Intra-session strategies: Just-in-time retrieval (dynamic file loading targeted fetching) context refresh mechanisms navigation and discovery during execution.

Post-session strategies: Summary generation (compression selective preservation) memory extraction and persistence formats enabling future continuity.

Implementation and evaluation:

• Design and implement 2-3 strategies at each lifecycle stage
• Benchmark on realistic multi-session development tasks
• Measure continuity quality token efficiency and coherence
• Analyze which combinations work best under different conditions

Possible extensions (if time permits):

• Investigate sub-agent architectures for parallel context handling
• Explore hybrid strategies combining pre-computed and just-in-time approaches
• Analyze how findings generalize across different types of agentic tasks

You will shape the research direction based on your discoveries and interests during the project.

Qualifications :

In order to be successful in the project:

• Master student(s) in Computer Science Software Engineering AI/Machine Learning or Data Science
• Passionate about Generative AI large language models (LLMs) and surrounding technologies
• Experienced with Python programming and comfortable working with APIs
• Interested in software engineering practices and agent-based systems
• Analytical with strong problem-solving skills and an interest in empirical research
• Self-driven and able to work independently while collaborating with the team
• Curious about emerging AI technologies and eager to contribute to cutting-edge research

Additional Information :

Supervisors: Samuel Peltomaa Staffan Lindgren

Scope: We encourage to have a team of 2 master thesis students working on the thesis.

Remote Work :

No

Employment Type :

Contract