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

The Problem: Collaboration Burden

Traditional pull requests (PRs) assume that the code diff is the primary artifact for collaboration. However, as AI agents increasingly generate code, the bottleneck has shifted from writing code to managing the collaboration burden. Reviewers must now decipher the intent behind agent-generated patches, verify their alignment with project policies, and assess long-term maintenance risks. When work crosses a trust boundary—such as from an external contributor, vendor, or contractor—the receiving project cannot simply trust the external implementation, yet the current PR interface forces them to review that external code as the primary merge candidate.

The KPR Workflow

Knowledge-Based Pull Requests (KPR) propose a fundamental shift: treat external code as a knowledge source rather than a merge candidate. In this workflow, an external collaborator uses their local agent to explore, test, and refine a solution. Instead of submitting the resulting diff, they submit a 'KPR package.' This package includes cleaned interaction traces, test results, and evidence of the problem-solving process.

This package is then processed through a collaboration gateway where agents distill the information into human-readable formats like design memos or risk checklists. Once a project-side human approves the knowledge, a project-owned 'inner trusted coding agent' regenerates the implementation within the project's own environment. This ensures the final code adheres to the project's specific engineering conventions, security policies, and architectural standards.

KPR separates two critical decisions that are currently collapsed in traditional PRs: whether the proposed knowledge should be accepted into the project, and whether a specific implementation is correct. By decoupling these, KPR allows projects to maintain strict governance over their codebase while still benefiting from the speed and exploration capabilities of external AI agents. This approach addresses the growing tension between open collaboration and the need for rigorous project-side control in an era of automated software development.