A Gated Spec-Driven Workflow: An Audit-Trail Study of Coding-Agent Work
Story 10, from arXiv
A study of a spec-driven coding workflow using Claude Code shows that structured, gated processes can achieve high task completion with minimal undeclared changes, though the orchestrator has notable limitations.
Key points
- The workflow uses a coding agent that follows specifications, plans, and task lists rather than chat transcripts.
- Each stage requires user approval, with a reviewer subagent checking for defects and an executor subagent limited to declared files.
- The orchestrator runs file-inventory checks to detect undeclared changes after each task block.
- Five out of seven projects completed all planned tasks, with 288 of 581 tasks completed block by block.
- The executor adhered to declared files, with only one unde, but the orchestrator failed to catch undeclared changes in 11 of 247 checks.
- The orchestrator stopped work in only four cases, often due to external factors like user requests or environmental changes.
Why it matters
This study provides insight into how structured, spec-driven workflows can improve reliability and traceability in AI-assisted coding, which is valuable for engineers looking to adopt more controlled and auditable practices.
Counterpoint
The orchestrator's reliance on self-checks introduces a bias, as it may not reliably detect changes caused by external factors or user interventions.
Sources
This summary was written by a language model and may contain errors. Read the original for the full story.