- The paper reveals that Collaborator tools lead to ≥96% agent-initiated PRs while nearly all merge decisions remain under human control.
- It employs a six-scenario taxonomy over 29,585 PRs to systematically quantify workflow transitions and review durations.
- Findings highlight practical implications for review workload planning and the need for richer live governance metadata integration.
Partitioning Operational Agency and Merge Governance in AI-Assisted PR Lifecycles
Introduction
The paper "Collaborator or Assistant? How AI Coding Agents Partition Work Across Pull Request Lifecycles" (2605.08017) presents an empirical analysis of the work partitioning and governance dynamics in software development workflows integrating AI coding agents. By disambiguating operational agency (who initiates PRs) from governance authority (who approves/merges), the authors challenge the conventional conflation of tool affordances with actual workflow practices. Using a taxonomy spanning six Initiator x Approver interaction scenarios, the study quantifies and structures the behavioral paradigms—Collaborator and Assistant—employed across five representative AI coding tools (OpenAI, Copilot, Devin, Cursor, Claude Code).
Analytical Framework
The lifecycle model defines PR trajectories in three non-terminal phases—Created, Review, Revision—and two terminal outcomes (Merged & Closed, Unmerged & Closed). Actor classification hinges on systematic bot-pattern heuristics mapped over 29,585 PRs from the AIDev dataset, yielding deterministic state machines per tool. The six-scenario taxonomy enables explicit separation of operational agency and governance authority, with Collaborator tools characterized by agent-led initiation and Assistant tools by human-led task direction.
Empirical Findings
Collaborator-Assistant Spectrum
Quantitative results demonstrate a strong association (Cramér's V = 0.50) between tool identity and interaction scenario. Collaborator-class tools (Cursor, Devin, Copilot) concentrate operational initiative in agents—≥96% agent-initiated PRs—while Assistant-class tools (OpenAI, Claude Code) exhibit ≥95.6% human-initiated PRs. Despite high agent operational engagement in Collaborators, terminal merge authority remains almost exclusively human (<0.1% agent-authorized merges). The decoupling is robust: agents do the operational work, but merge governance is rarely delegated.
Workflow Dynamics
Collaborator tools systematically route PRs through Review, with substantive revision loops (Copilot: 90.3%, Cursor: 51.3%, Devin: 52.2% into Review; Copilot median Review time: 3.0 hours). Assistant tool workflows, particularly OpenAI, resolve the majority of PRs directly from development to merge (OpenAI: 76.5% direct resolution, median Review time: 0.7 hours). Controlling for initiation shows the review-routing gap persists even among human-initiated PRs, indicating PR phase transitions are influenced by tool paradigm but also by repository governance and adoption context.
Automation-Authorized Merge Analysis
Agent-classified approvers (S2/S5) comprise a negligible fraction (0.24% of merged PRs), with most automation merges resulting from infrastructure or repository policy. Event logs capture execution, not governance decision; only live repository metadata can resolve the boundary between mechanism and authority. The empirical upper bound of observed autonomous agent governance is 0.07%—a strong indication of contemporary organizational conservatism toward delegating merge authority to agents.
Implications
Practical
- Review Workload Planning: Collaborator tools impose significant human review overhead, with deterministic sojourn times and transition probabilities usable for estimation in sprint planning.
- Risk-Based Review: Assistant workflows facilitate direct resolution, suggesting suitability for graduated trust models—mandating review for complex/high-risk PRs only.
- Tool Selection and Governance: The taxonomy provides actionable guidance for team-level review allocation and organizational governance policy.
Theoretical
- Agency-Governance Decoupling: The study substantiates the empirical separation between operational agency and governance authority, informing theoretical work on human-AI delegation and distributed cognition.
- Boundary of Observation: The measurement boundary exposed by automation-authorized merges signals the need for richer governance metadata; event logs alone cannot establish decision provenance.
Future Directions
- Integration of Live Governance Metadata: Resolution of merge authority questions necessitates live API access to branch protection rules, required reviewers, and policy configurations.
- Quality-Effort Correlation: Linking interaction scenarios with PR complexity, test coverage, and post-merge defect rates would clarify potential tradeoffs in direct resolution versus review-intensive workflows.
- Review Content Analysis: Differentiating substantive from superficial reviews will require content-level analysis, potentially integrating NLP models for comment depth.
Conclusion
The study operationalizes and empirically maps the Collaborator-Assistant spectrum in AI-assisted software development PR lifecycles, revealing systematic agent-led initiation but persistent human-centric merge governance. The introduced taxonomy and workflow state machines present a structured quantitative baseline for further research and practical tool adoption. Future work must bridge the measurement boundary of merge authority and explore quality-governance tradeoffs as agentic coding tool prevalence grows.
