Building GitHub Actions-based LLM agents — Leverage the open-source Agentic Workflows templates to compose multi-step, goal-driven automation directly in GitHub Actions with natural-language task definitions.
Implementing prompt-injection defenses — Integrate the integrity filtering module into any agentic pipeline that processes untrusted repository content (issues, PR comments, commit messages) before passing data to an LLM.
Automated codebase migrations with verification — Use Crane's agentic planner to break large-scale refactors (e.g., API version upgrades, dependency swaps) into atomic, testable steps with rollback support.
LLM-powered repository indexing and search — Prototype semantic codebase navigation using Repo Mind / Copernicus as a reference architecture for embedding and querying large codebases.
Modular agent prompt engineering — Apply the Agent Functions pattern to decompose monolithic agent prompts into reusable, parameterized subroutines for cleaner, more maintainable LLM pipelines.
Research into continuous AI in software platforms — Study the Continuous AI project as a reference for integrating persistent, event-driven LLM automation into collaborative software development platforms.

GitHub Next