# Designing Human-Agent Collaboration Workflows: Research → Plan → Execute → Review → Ship
AI4SE process design isn't about inserting AI into an old pipeline — it's about redesigning the end-to-end workflow around 'who makes which decisions, and what the agent executes.'
The AI4SE Evolution of Classic Workflows
Traditional software engineering processes are designed human-first. The core question in the AI4SE era is: who leads each step?
| Step | Primary role | Agent’s role | Key decision point |
|---|---|---|---|
| Research | Human + Agent | Explore, search, draft | Direction confirmation |
| Plan | Human-led | Generate solution options | Spec sign-off |
| Execute | Agent | Sandboxed execution | Risk routing |
| Review | Human + independent Evaluator | Cannot self-assess | Quality gate |
| Ship | Human accountable | Automated deployment | Merge approval |
Three Principles for Process Design
- The decision-maker doesn’t execute: the person who writes the spec doesn’t review the code produced against that same spec
- Risk routing: low-risk actions are agent-autonomous, high-risk actions require HITL
- Complete evidence chain: the Spec, Diff, and Log produced at every step form an auditable record
Anti-Patterns
- Treating the agent as a search engine — only asking, never changing anything
- Skipping review and merging directly — amplifies the agent’s blind spots
- Requiring human approval at every step — process cost eats away AI’s efficiency gains