openclaw-model-orchestrator_skill

Overview

This skill orchestrates multi-LLM workflows for OpenClaw, enabling fan-out, pipeline, and consensus patterns across 40+ models. It uses AAHP v3–style structured handoffs to minimize token overhead while routing tasks between planner, worker, and reviewer roles. The skill includes smart recommendations and pre-configured task profiles for common engineering and research tasks.

How this skill works

The orchestrator decomposes incoming tasks and dispatches work according to the chosen mode: fan-out for parallel subtasks, pipeline for sequential refinement, and consensus for aggregated answers. It encodes model-to-model communication in compact AAHP v3 handoff objects that contain context, routing metadata, differential state, and constraints. A planner generates the plan, workers execute subtasks, and a reviewer merges or synthesizes results; an automatic recommender suggests optimal model combinations.

When to use it

• Parallelize independent subtasks to speed up large jobs (fan-out).
• Chain specialized models for staged workflows like plan → implement → review (pipeline).
• Aggregate multiple model opinions for robust answers or disputed topics (consensus).
• Auto-recommend model combos for unfamiliar tasks or to optimize cost/quality trade-offs.
• Run mass or bulk operations using pre-configured task profiles.

Best practices

• Start with a task profile (coding, research, security, review, bulk) to use proven model combinations.
• Use the recommender when unsure which planner/workers/reviewer to pick.
• Constrain outputs explicitly in handoff constraints to reduce ambiguity and token usage.
• Limit maxConcurrent workers to control parallelism and downstream load.
• Review differential state in handoffs for efficient debugging and incremental runs.

Example use cases

• Build a REST API: planner breaks features into endpoints; workers implement endpoints in parallel; reviewer integrates and tests.
• Security audit: planner defines threat categories; security-specialist workers analyze each area; reviewer synthesizes vulnerabilities and fixes.
• Research synthesis: send same query to multiple research models and synthesize consensus plus unique insights.
• Feature rollout: pipeline through plan, implement, integration test, and polish stages with specialist models at each step.
• Bulk transformations: apply a standardized transformation across thousands of documents using the bulk profile and parallel workers.