Skip to content

Agents

Vandelay agents are built on Agno's Agent class with additional layers for workspace-driven prompts, hot-reloadable tools, and persistent memory.

Agent Lifecycle

  1. Config loaded: Settings read from config.json, .env, and environment variables
  2. Model created: LLM provider initialized from model.provider and model_id
  3. Tools resolved: Enabled tools instantiated from the tool registry
  4. Workspace assembled: Seven markdown files compose the system prompt
  5. Agent created: Agno Agent (or Team) instantiated with all components
  6. Ready: Agent accepts messages via ChatService

System Prompt Assembly

The system prompt is built from workspace files in ~/.vandelay/workspace/:

File Purpose
SOUL.md Core personality, communication style, behavioral rules
USER.md User preferences, context, personal details
AGENTS.md Team delegation rules: which member handles what
TOOLS.md Tool usage guidance, best practices per tool
HEARTBEAT.md Proactive tasks to check during heartbeat
BOOTSTRAP.md One-time startup instructions
MEMORY.md Persistent notes the agent maintains

These files are yours to edit. Changes take effect on the next agent reload.

Hot Reload

When you enable or disable a tool (via CLI or chat), the agent is recreated with the new tool set without a server restart. The AgentProvider protocol ensures ChatService always gets the current agent instance.

Single Agent vs Team

  • Single agent (team disabled): One agent with all enabled tools
  • Team mode (default): A supervisor delegates to specialist members with scoped tools

See Teams for details on team mode.