NHN Cloud Tech Blog Shares 'Artifact Driven Development' Experience for Parallel AI Agent Development

To address the bottleneck caused by developers manually monitoring execution processes when running AI coding agents in parallel, the blog proposes a development methodology called Artifact Driven Development. This approach judges results by reviewing only the final deliverables—Plan, Test Case, and Code—without observing the agents' intermediate processes.

The process-observation approach lacks scalability in parallel tasks involving three or more agents, and observing intermediate processes does not guarantee result quality. In contrast, deliverable-centric review is scalable regardless of the number of agents and minimizes developer intervention.

Deliverable-Centric Development Structure

This methodology is based on three core deliverables. Plan defines the implementation scope and file modifications to constrain the agent's direction and serves as a mechanism for context recovery at session end. Test Case (TC) acts as an objective standard for verifying implementation completion, guiding agents to automatically run and fix tests through automation. Code represents the final changes that pass TC within the Plan scope, verified through diff review.

Orchestrator-Executor Pattern and Environment Isolation

To efficiently coordinate multiple agents, the Orchestrator-Executor pattern was applied. The Orchestrator handles discussions with developers and task dispatching, with code modification prohibited. Executors start with a new session each time to enforce the precision of the Plan. Additionally, Git Worktree is used to separate code directories, and server ports, DB, and build caches are isolated by feature to block interference during parallel execution.