Thesis: Context Engineering for Autonomous Software Development: Bridging the Model Capability Gap
Knightec Group ABSammanfattning
This thesis position focuses on developing a reusable context engineering pipeline for autonomous AI coding agents, aimed at improving the performance of smaller language models in software development tasks. The project will be conducted at Knightec Group's office in Sundsvall, starting in January or March 2027. The research will evaluate various context engineering strategies and their effectiveness in bridging the performance gap between smaller and frontier models.Jobbet i korthet
Anställningstyp
studentjobb
Det här erbjuder vi
Opportunity to conduct research in a cutting-edge field of AI and software development.Work in a collaborative environment at a leading strategic partner in product and digital service development.Gain practical experience in context engineering and AI coding agents.
Ansök senast: Öppet tillsvidare
Publicerad: 2026-10-15
Beskrivning
High Level Description
Autonomous AI coding agents increasingly rely on powerful frontier language models to understand and modify complex software systems. However, these models can be expensive to operate, while smaller and more affordable models often struggle with the same development tasks.
Context engineering offers a potential way to narrow this gap by systematically selecting, structuring, and supplying relevant information about the codebase, architecture, development conventions, and task at hand.
This thesis investigates whether well-engineered context can enable smaller or less capable language models to approach the performance of frontier models on realistic software development tasks. It also explores how much context is actually needed, and whether providing too much information can reduce performance rather than improve it.
Project Description
In this thesis, investigate and build a reusable context engineering pipeline for autonomous AI coding agents. The pipeline should support different strategies for selecting, structuring, and supplying context to language models working on software development tasks within an existing codebase.
Three context engineering approaches should be compared:
Persistent project context: Human-authored documentation, architecture descriptions, coding guidelines, and project instructions that remain available across tasks.
Task-specific manual context: Relevant code, dependencies, and architectural information selected and prepared by a human for a particular development task.
Automatically generated context: Relevant information identified and assembled automatically through repository analysis, code search, retrieval, or summarization.
These approaches should be evaluated against a baseline coding agent operating without additional prepared context, while retaining its normal ability to explore the repository and use development tools.
The evaluation should compare smaller and frontier language models performing the same tasks in representative industrial-style software repositories. The solution should be model-agnostic, supporting both local and cloud-based inference. Local execution of smaller models is optional, not a requirement.
The prototype should preferably be built on an existing autonomous coding agent harness, such as OpenCode, rather than requiring a new agent framework.
Particular attention should be given to context quantity and relevance. By systematically varying the amount of useful information and introducing redundant or unrelated context, investigate when additional context stops improving performance or begins to negatively affect the agent's results.
The evaluation should use reproducible development tasks, such as feature implementation, bug fixing, and refactoring, with automated tests and clearly defined success criteria. It should also distinguish between context prepared before task execution and information the agent discovers independently through repository exploration.
Who are we looking for?
Purpose
I n this thesis, investigate these questions:
The primary evaluation should measure functional task success and the reduction in the performance gap between smaller and frontier models, without requiring a predefined equivalence threshold.
Experiments should use consistent agent tooling, reproducible repository states, and comparable execution conditions.
Out of scope: Training or fine-tuning language models, developing a new autonomous coding agent framework, extensive hardware benchmarking, optimizing local inference infrastructure, and establishing universally optimal context configurations across all software development environments.
References
Practical Information
This is a thesis position, located at our office in Sundsvall . Start date January or March 2027.
Please submit your application as soon as possible, but no later than 2026-11-30 . If you have any questions, you are welcome to contact johanna.edstrom@knightec.se . Note that due to GDPR, we only accept applications through our careers page.
Autonomous AI coding agents increasingly rely on powerful frontier language models to understand and modify complex software systems. However, these models can be expensive to operate, while smaller and more affordable models often struggle with the same development tasks.
Context engineering offers a potential way to narrow this gap by systematically selecting, structuring, and supplying relevant information about the codebase, architecture, development conventions, and task at hand.
This thesis investigates whether well-engineered context can enable smaller or less capable language models to approach the performance of frontier models on realistic software development tasks. It also explores how much context is actually needed, and whether providing too much information can reduce performance rather than improve it.
Project Description
In this thesis, investigate and build a reusable context engineering pipeline for autonomous AI coding agents. The pipeline should support different strategies for selecting, structuring, and supplying context to language models working on software development tasks within an existing codebase.
Three context engineering approaches should be compared:
Persistent project context: Human-authored documentation, architecture descriptions, coding guidelines, and project instructions that remain available across tasks.
Task-specific manual context: Relevant code, dependencies, and architectural information selected and prepared by a human for a particular development task.
Automatically generated context: Relevant information identified and assembled automatically through repository analysis, code search, retrieval, or summarization.
These approaches should be evaluated against a baseline coding agent operating without additional prepared context, while retaining its normal ability to explore the repository and use development tools.
The evaluation should compare smaller and frontier language models performing the same tasks in representative industrial-style software repositories. The solution should be model-agnostic, supporting both local and cloud-based inference. Local execution of smaller models is optional, not a requirement.
The prototype should preferably be built on an existing autonomous coding agent harness, such as OpenCode, rather than requiring a new agent framework.
Particular attention should be given to context quantity and relevance. By systematically varying the amount of useful information and introducing redundant or unrelated context, investigate when additional context stops improving performance or begins to negatively affect the agent's results.
The evaluation should use reproducible development tasks, such as feature implementation, bug fixing, and refactoring, with automated tests and clearly defined success criteria. It should also distinguish between context prepared before task execution and information the agent discovers independently through repository exploration.
Who are we looking for?
- Master of Science students in Computer Science, Software Engineering, or related fields.
- Graduation in June 2027
- Fluency in Swedish, both written and spoken
- Fluency in English, both written and spoken
Purpose
I n this thesis, investigate these questions:
- Can context engineering compensate for lower model capability?
- How much can engineered context reduce the performance gap between smaller language models and frontier models?
- For which types of software development tasks is context engineering most effective?
- Which context engineering strategies work best?
- How do persistent project documentation, manually curated task-specific context, and automatically generated context compare?
- Does their effectiveness differ depending on model capability?
- To what extent does prepared context reduce the need for independent repository exploration?
- How much context is enough?
- How does the amount of relevant context influence task completion performance?
- At what point does additional context provide diminishing returns?
- Can excessive, redundant, or irrelevant context negatively affect performance, particularly for smaller models?
- What are the practical trade-offs?
- How do different context strategies affect task success, execution time, inference cost, and context preparation effort?
- Can smaller models with engineered context provide a more cost-effective alternative to frontier models?
- How complex is it to implement and maintain a reusable context engineering pipeline?
The primary evaluation should measure functional task success and the reduction in the performance gap between smaller and frontier models, without requiring a predefined equivalence threshold.
Experiments should use consistent agent tooling, reproducible repository states, and comparable execution conditions.
Out of scope: Training or fine-tuning language models, developing a new autonomous coding agent framework, extensive hardware benchmarking, optimizing local inference infrastructure, and establishing universally optimal context configurations across all software development environments.
References
- OpenCode - https://opencode.ai/
- SWE-bench - https://www.swebench.com/
- Anthropic, Effective Context Engineering for AI Agents - https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
An Exciting Journey with Knightec Group
Semcon and Knightec have joined forces as Knightec Group. Together, we are Northern Europe's leading strategic partner in product and digital service development. With a unique combination of cross-functional expertise and a holistic business understanding, we help our clients realize their strategies - from idea to complete solution.
Practical Information
This is a thesis position, located at our office in Sundsvall . Start date January or March 2027.
Please submit your application as soon as possible, but no later than 2026-11-30 . If you have any questions, you are welcome to contact johanna.edstrom@knightec.se . Note that due to GDPR, we only accept applications through our careers page.
Ansök till tjänsten
Thesis: Context Engineering for Autonomous Software Development: Bridging the Model Capability Gap
OM FÖRETAGET
Knightec Group AB
