feat: Agent-to-agent handoff and delegation

[ecschoye]

Summary

Enable multi-agent workflows where a primary agent can delegate subtasks to specialized agents with different models or configurations.

Motivation

Complex tasks benefit from specialization:

• Use a fast/cheap model for simple subtasks (summarization, classification)
• Route reasoning-heavy tasks to a stronger model
• Parallel execution of independent subtasks
• Separation of concerns (research agent vs coding agent)

Requirements

• delegate tool: spawn a sub-agent with a specific prompt and model
• Model routing: delegate to different provider/model combinations
• Context passing: share relevant context with sub-agents
• Result aggregation: collect and synthesize sub-agent outputs
• Recursion limits (prevent infinite delegation loops)
• Cost tracking per sub-agent

Configuration

[agents.researcher]
provider = "gemini"
model = "gemini-2.0-flash"
system_prompt = "You are a research assistant..."

[agents.coder]
provider = "openrouter"
model = "anthropic/claude-sonnet-4-20250514"
system_prompt = "You are a coding assistant..."

Implementation Notes

• Build on existing RouterProvider hint system
• Sub-agents share the same tool registry
• Consider async execution with result futures
• Leverage ReliableProvider for sub-agent calls

🤖 Generated with Claude Code