OpenClaw Model Orchestrator

Dispatch tasks across multiple LLMs from chat. Uses AAHP v3 structured handoffs for minimal token overhead.

Orchestration Modes

Fan-Out

Split a task into parallel subtasks, each executed by a different model.

A planner model decomposes the task, workers execute in parallel, a reviewer merges results.

/orchestrate --mode fan-out --task "Build a REST API with auth" --planner copilot-opus --workers copilot52c,grokfast --reviewer copilot-sonnet46

Pipeline

Chain models sequentially. Each model refines the previous model's output.

Ideal for plan -> implement -> review -> polish workflows.

/orchestrate --mode pipeline --task "Design and implement a caching layer" --planner copilot-opus --workers copilot52c,copilot-sonnet46 --reviewer copilot-opus

Consensus

Send the same question to multiple models, then synthesize the best answer.

Identifies agreement, disagreement, and unique insights across models.

/orchestrate --mode consensus --task "What are the security risks of this API design?" --workers copilot-opus,gemini25,sonnet --reviewer copilot-opus

Smart Recommendations

The orchestrator auto-classifies tasks and recommends optimal model combinations:

/orchestrate recommend "Build a REST API with JWT auth"

Returns: task classification, recommended planner/workers/reviewer, reasoning, and a ready-to-run command.

Use help as any flag value for context-aware recommendations:

/orchestrate --task "Audit security" --planner help

Task Profiles

Pre-configured model combinations optimized for common task types:

AAHP v3 Integration

All model-to-model communication uses structured AAHP v3 handoff objects instead of raw chat history. This achieves up to 98% token reduction compared to naive context passing. Each handoff contains:

• Task context (only relevant information)
• Routing metadata (source/target model, mode)
• Differential state (only what changed)
• Constraints (output format, scope limits)

Commands

Configuration

In openclaw.plugin.json:

{
  "config": {
    "defaultPlanner": "copilot-opus",
    "defaultReviewer": "copilot-sonnet46",
    "defaultWorkers": ["copilot52c", "grokfast", "copilot51"],
    "maxConcurrent": 4,
    "taskProfiles": { ... }
  }
}