vstack: A New Paradigm for Building Structured Multi-Agent Workflows in VS Code

vstack is an AI engineering workflow system natively integrated into VS Code, designed to address the pain points of multi-agent collaboration in AI-assisted programming. It covers the complete software development lifecycle from requirements to release through six fixed roles—Product, Architecture, Design, Engineering, Testing, and Release—and uses a template-driven workflow model to enable AI agents to collaborate like a team.

Today, with the popularity of AI-assisted programming tools, developers face the challenge of multi-agents working in silos. Inspired by gstack, vstack was rebuilt entirely around the native VS Code environment and uses a template-driven model. Its core positioning is an AI engineering workflow system within VS Code, providing structured skills (executable by GitHub Copilot in Agent mode) covering planning, review, verification, release, and other stages, adapting to product visions through fixed roles.

• Product (@product)：Entry point for requirements, responsible for requirement collection and acceptance criteria definition. Skills include vision, requirements, etc. Default mode: compact.
• Architecture (@architect)：Technical design review and decision recording. Skills: architecture, adr. Default mode: standard.
• Design (@designer)：API contracts and interface specifications. Skills: design, openapi, etc. Default mode: compact.
• Engineering (@engineer)：Core of code implementation. Skills: code-review, debug, etc. Default mode: compact.
• Testing (@tester)：Quality assurance. Skills: verify, security, etc. Default mode: ultra-compact.
• Release (@release)：Release gatekeeping and release note generation. Skills: release-notes, pr, etc. Default mode: compact.

vstack supports powerful workflow orchestration: The Product agent can call other roles as sub-agents and schedule them according to dependency relationships (e.g., Requirements → Architecture → Design → Engineering → Testing → Release). Skill invocation is flexible: directly @role name or /skill name (e.g., /verify). Response density can be adjusted via /concise: normal (full explanation), compact (brief and accurate), ultra (facts only/commands), status (show current settings). Security warnings and destructive operation prompts always use the normal mode.

Model selection recommendations: For deep reasoning tasks (product planning, architecture review), Claude Sonnet/Opus 4.6 is recommended; for testing, engineering, and release tasks, use Claude Sonnet 4.6; for complex debugging, upgrade to Claude Opus or GPT-5.3 Codex. Typical new feature workflow: /vision to verify feasibility → /architecture to lock design → coding → /verify for testing → /release for deployment. AI acts as a collaborative partner rather than a replacement for developers.

vstack is developed with Python 3.11-3.14 and installed via pip. It can be installed to a specific project (vstack install --target) or global VS Code configuration (--global). It uses a template-driven architecture, with templates located in src/vstack/_templates/, including categories like skills and agents. Each template consists of config.yaml (metadata) and template.md (instructions), making it easy to extend new roles or skills.

vstack solves the fragmentation problem of AI programming tools and promotes the evolution from single-point tools to systematic workflows. For teams, it provides repeatable and auditable processes, improving development efficiency, software quality, and collaboration predictability. In the future, such systematic frameworks will become standard configurations for large-scale projects, helping developers better leverage AI to build complex and reliable systems.