Turning Gemini CLI into a Multi-Agent System with Just Prompts (aipositive.substack.com)

A developer demonstrated how to turn Google’s Gemini CLI into a lightweight multi-agent orchestration system using only native features and prompt engineering—no extra code. By authoring custom commands (.toml files) and adopting a filesystem-as-state pattern (.gemini/agents/{tasks,plans,logs,workspace}), the orchestrator writes task JSON files and launches independent gemini-cli processes (each loaded with a single extension) as stateless workers. Each /agents:run invocation constructs a shell command like gemini -e <agent> -y -p "<prompt>" to spin up a sub-agent that reads its task, acts, writes logs, and exits. The whole system is transparent, debuggable, and leverages Gemini CLI’s extension tooling and prompt-driven behavior. The experiment highlights practical implications for AI/ML workflows: you can encode specialized expertise into isolated extensions (e.g., secure code reviewer, cloud provisioning) and orchestrate them via simple disk-based queues. Key technical lessons include avoiding recursive delegation by giving child processes an explicit first-person identity and Task ID in the prompt, and the trade-off between explicit agent assignment versus opaque automated selection. Important cautions: the -y/--yolo auto-approve flag can enable risky tool calls or rogue processes, so supervise runs and scope tools carefully. Overall, this shows prompt-native orchestration can already enable composable, enterprise-grade agent patterns without waiting for built-in background-agent features.