Orchestrate a multi-phase RAG migration
Posted on Apr 28 • Originally published at Medium. Task: The author orchestrated a multi-phase RAG migration using Gemini CLI and the Conductor extension. The workflow involved defining project specs in Markdown files (product, tech stack, tracks) to provide system-level context. The agent executed tasks via a Plan-Act-Validate cycle, generating Terraform infrastructure and Python/Next.js code. The human reviewed and approved spec.md and plan.md files before any implementation began, ensuring alignment with the technical strategy.; tools: Gemini. Human role: The original post describes the human role this way: Only after I review and approve them, the actual implementation starts.; limitations: The source does not explicitly state hard limitations, but the workflow relies on the Conductor extension and Gemini CLI. Human approval is mandatory before implementation, creating a bottleneck for speed. Token costs vary based on usage patterns, and the $30 figure is specific to this project's scale. Integration with existing CI/CD pipelines and handling mid-track spec evolution are not fully detailed..
Tools
Published on Agent Work Atlas:
Overview
Posted on Apr 28 • Originally published at Medium
Workflow
The author orchestrated a multi-phase RAG migration using Gemini CLI and the Conductor extension. The workflow involved defining project specs in Markdown files (product, tech stack, tracks) to provide system-level context. The agent executed tasks via a Plan-Act-Validate cycle, generating Terraform infrastructure and Python/Next.js code. The human reviewed and approved spec.md and plan.md files before any implementation began, ensuring alignment with the technical strategy.
Reported result
The project achieved a target of >80% test coverage for new code, forcing the agent to verify functionality rather than just generating syntactically correct code. The total Vertex AI token cost was approximately $30 for several days of work, with significant savings from cached input tokens. The workflow successfully managed infrastructure, backend, and frontend components while maintaining consistent technical decisions through spec-driven development.
Human role
The original post describes the human role this way: Only after I review and approve them, the actual implementation starts.
Limitations
The source does not explicitly state hard limitations, but the workflow relies on the Conductor extension and Gemini CLI. Human approval is mandatory before implementation, creating a bottleneck for speed. Token costs vary based on usage patterns, and the $30 figure is specific to this project's scale. Integration with existing CI/CD pipelines and handling mid-track spec evolution are not fully detailed.
Try it
I want to orchestrate a new cloud project using Gemini CLI and the Conductor extension. First, initialize the project by running /conductor:setup and defining product.md and tech-stack.md. Then, create a new track for [replaceable feature name] using /conductor:newTrack. The agent should generate spec.md and plan.md. I will review and approve these files before you run /conductor:implement. Ensure you follow the TDD protocol: write failing tests, implement code to pass them, and verify >80% coverage before committing. Use Terraform for infrastructure and run terraform plan before apply. Keep me in the loop for all approvals.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.