multi-agent-thinker_skill

Overview

This skill is the Heavy Lifting Engine for complex problem solving. It coordinates multiple lightweight agents to expand reasoning capacity, isolate context, and parallelize subtasks. The goal is reliable, scalable cognition for ambiguous, multi-step, or high-stakes requests.

How this skill works

The skill evaluates the user request for ambiguity, planning needs, multiple perspectives, or explicit cognitive-strategy triggers. When invoked, it spawns a coordinator and specialized sub-agents using one of three patterns (supervisor/orchestrator, peer-to-peer swarm, or hierarchical) and routes work with clear handoff protocols. It enforces context isolation, validation checkpoints, and consensus or debate protocols to mitigate hallucination and error propagation.

When to use it

• Request is ambiguous or underspecified (e.g., “Design a system…”).
• You need reasoning, architecture, or planning before coding.
• You want multiple perspectives, critique, or debate on a decision.
• You ask explicitly for cognitive strategies (Socratic, pre-mortem, devil’s advocate).
• The task requires parallel analysis of independent sources or documents.

Best practices

• Choose architecture pattern to match task: supervisor for strict control, swarm for exploration, hierarchical for layered workflows.
• Keep sub-agent contexts minimal: pass instructions or use persistent files instead of full context dumps when possible.
• Implement direct pass-through for final sub-agent outputs to avoid supervisor paraphrase errors.
• Weight agent votes by confidence and expertise; use adversarial critique rounds for high-stakes conclusions.
• Monitor behavioral triggers (stall, sycophancy, divergence) and apply circuit breakers or retries.

Example use cases

• Designing complex software architecture with trade-off analysis across modules.
• Running comparative research across many documents or sources in parallel.
• Pre-mortem risk analysis for a project with adversarial critique rounds.
• Reverse-engineering mature products to derive an inferred spec and feature set.
• Creating multi-layered workflows for enterprise processes requiring strategic, planning, and execution coordination.