Multi-Agent Collaboration

Status: Not Implemented

OpenViber currently uses a single-agent architecture. A single well-configured agent with skills is sufficient for most use cases. This document describes potential future multi-agent patterns if needed.

Current Architecture

OpenViber follows a single-agent model:

┌─────────────────────────────────────────────────────────────┐
│  Human (Manager)                                             │
│  "Build a landing page with a contact form"                  │
└─────────────────────────────────────────────────────────────┘
                                  │
                                  ▼
┌─────────────────────────────────────────────────────────────┐
│  Viber (Single Agent + Skills)                               │
│  Plans work, uses tools, executes with skill knowledge       │
└─────────────────────────────────────────────────────────────┘

Why Single Agent?

For the stateless daemon architecture:

Simplicity: One agent with appropriate skills handles most tasks
Stateless: No coordination state to maintain between agents
Context efficiency: Single agent means single context window
Sufficient capability: Skills provide domain knowledge; tools provide actions

When Multi-Agent Might Be Needed

Scenario	Single Agent Solution	When Multi-Agent Helps
Different domains	Add skills for each domain	Domains require conflicting behaviors
Parallel work	Sequential execution	True parallelism on independent subtasks
Review/QA	Self-review prompts	Formal separation of concerns

Future Multi-Agent Patterns

If multi-agent is implemented in the future:

Pattern 1: Specialist Delegation

Viber (Orchestrator)
  ├── Developer Agent (coding skill)
  ├── Reviewer Agent (review skill)
  └── Designer Agent (design skill)

Pattern 2: Pipeline

Planner → Executor → Verifier → Reporter

Implementation Considerations

Stateless coordination: All state in Viber Board, not daemon
Context passing: How to share context between agents efficiently
Cost: Multiple agents = multiple LLM calls
Complexity: Is it worth the added complexity?

Recommendation

For most use cases, stick with a single agent configured with:

Clear system prompt defining the role
Appropriate skills for domain knowledge
Required tools for actions
Good working mode (Always Ask / Agent Decides / Always Execute)

Multi-agent should only be considered when there’s a clear benefit that can’t be achieved with skills and tools.