Knowledge-Based Pull Requests: A Trusted Workflow for AI Agent Code Collaboration

The paper "Knowledge-Based Pull Requests" from arXiv (2606.26721) addresses a critical problem in AI-assisted software development: while AI coding agents make code generation cheap, understanding intent, negotiating scope, and governing long-term responsibility remain costly. The authors propose KPR, a workflow where external collaborator code, tests, and agent interaction traces are treated as knowledge sources — not direct merge candidates. An agent distills these into a human-confirmed knowledge package (design memo, risk checklist, test plan), then a project-owned inner trusted agent regenerates candidate code inside the receiving project's environment, respecting conventions, tests, and security policies.

The key insight is separating two decisions: whether the knowledge should enter the project, and whether a particular implementation should be merged. The paper contributes a candidate artifact schema, cost-accounting view, collaboration gateway architecture, and a minimal controlled simulation pilot over seven merged public pull requests. The pilot shows KPR packages can be instantiated from real PR material and stress-tested under description ablation, diff ablation, and synthetic poisoned-patch conditions. The work positions KPR as an empirically testable workflow whose value depends on whether auditable extraction and regeneration reduce the cost of understanding high-context external changes.