Lieutenant workflow: coordination-only agent

Design an agent role that ONLY coordinates subagents - no direct coding/execution.

Concept: The Lieutenant

A pure orchestration agent that:

• Receives high-level goals
• Decomposes into subagent tasks
• Monitors subagent progress
• Integrates results back to main branch
• Never writes code or runs tools directly (except spawn/monitor subagents)

Why?

1. Separation of concerns: Coordination logic separate from execution 2. Scalability: Can manage many parallel subagents 3. Reliability: Not polluted by execution context; clean decision-making 4. Integration responsibility: Owns the merge/integration step

Responsibilities

• Task decomposition and assignment
• Subagent health monitoring
• Conflict resolution between subagent outputs
• Integration: merge subagent branches, resolve conflicts
• Escalation: know when to ask human

Tools

• spawn_subagent(task, config) - start a worker agent
• check_subagent(id) - get status/progress
• kill_subagent(id) - cancel runaway agent
• integrate_work(subagent_ids) - merge their work to main
• NO: read_file, bash, edit, etc.

Integration with Existing System

• Extends current orchestrator in Omni/Ava.hs
• Subagents are coder/researcher agents
• Lieutenant handles git branch management

Open Questions

1. How does lieutenant evaluate subagent output quality? 2. What's the handoff protocol? (structured output? branch state?) 3. How to handle subagent that needs more info from lieutenant?